08自由意志 Peter W. Kalivas, Martin P. Paulus - Intrusive Thinking - From Molecules to Free Will-MIT Press (2021)

The Day Sue Almost
Got Mad at Me
Have you ever wondered why thoughts appear and can motivate you to do
things you wouldn’t otherwise do? One afternoon, Sue (my wife) called me at
work to tell me she was going out with friends after work, and it was up to me
to get dinner and do homework with the kids. Sue and her messages were (and
remain to this day) very motivating for me, so I packed up and went down to
my car. On the way to the car and during the drive home, I randomly thought
about what I would cook for dinner, which kid was having a test and would
need help, and various other plans for the evening. About halfway home,
Kyle from the pub called. He was energized because a buddy from southern
California had just flown in and thus Kyle was rounding people up for what
sounded like a really fun night. I was going to tell him, “Sorry, I can’t make it,”
but then the 20 beers on tap suddenly popped into my mind, as did the hilari-
ous time we had the last time we all got together. Pretty soon I wasn’t thinking
about dinner plans or school tests any more. I was making plans to be with my
friends and thinking about the latest release IPA from the local microbrewery.
When we ended our call, I had told him that I wasn’t sure what I would do.
At this point, we might ask ourselves what is more motivating: going home to
cooking and homework, or going to the pub for beer and friends. Each one of
us faces these types of decisions, big and small, all the time, and what we do
depends largely on how we organize our thoughts about each possible scenario
and which ultimately seems most important to us. In this story, if you know
Sue, you pretty much know already that I went home and had to hear about
all of the pub hilarity the next morning. But, if I suffered from alcohol use
disorder, thoughts about what would happen at the pub, the taste of the beer,
and past really fun pub experiences would have inevitably intruded until they
all but squeezed out thoughts of dinner and homework with my kids. My plans
to go home would begin to fade until they were all but forgotten, or perhaps
I would rationalize that stopping by for one of the special IPAs before going
home would somehow work out.
The story above illustrates how thoughts of pub friends and beer can intrude
in substance use disorder. However, the intrusion of traumatic events in post-
traumatic stress disorder, rumination on negative outcomes in depression, or
hearing voices in schizophrenia are all examples of thoughts generated by your
brain that can contribute to debilitating psychiatric disorders. Of course, it is a
natural and healthy adaptive process to produce thoughts either randomly or in
From “Intrusive Thinking: From Molecules to Free Will,” edited by Peter W. Kalivas and Martin P. Paulus.
Strüngmann Forum Reports, vol. 30, Julia R. Lupp, series editor. Cambridge, MA: MIT Press. ISBN 978-0-262-54237-1
xvi P. W. Kalivas
association with the world we are experiencing, and then to use these thoughts
to navigate successfully toward desired outcomes.
This volume explores and provides the best possible explanations for what
this process is, how it gets usurped in psychiatric disorders, and what this
knowledge of how the brain handles thoughts means for concepts of free will
and one’s responsibility for poor decisions, especially when a thought disorder
exists. It addresses how the brain is organized to create thoughts that can be ig-
nored or can build in motivational content, and how we then weigh thoughts to
decide on behavior that best adapts us to the world. It also poses and attempts
to answer a number of questions that are commonly asked: How do the mecha-
nisms of thought intrusion and decision making get corrupted in psychiatric
disorders to create intrusions that cannot be controlled? How are thought intru-
sions usurped by motivation to produce behavior that may be maladaptive, at
least according to social norms? What is free will and what responsibility does
free will (or lack of it) create for how we behave?
— Peter W. Kalivas
Molecules and Circuits
2
What Are the Circuits

That Mediate and Update
Intrusive Thinking?
Paul E. M. Phillips and Amy L. Milton
Abstract
This chapter discusses psychological constructs considered to be central to the media-
tion of intrusive thinking and the neural circuits that underlie these processes. It assimi-
lates intrusive thoughts with conditioned responses, discerns associate structures that
can support these responses, and suggests how episodic information may be integrated
with these associations. Mechanisms by which intrusive thoughts can be updated are
explored, with a focus on extinction and memory reconsolidation. Intrusive thoughts
ultimately engage many areas of the brain as they encompass sensory, cognitive, mo-
tor, and somatic processes. In this chapter, the focus is on specific circuits within the
prefrontal-limbic network that are proposed to encode, update, and maintain the con-
tent of intrusions. These circuits include interconnecting pathways between the ventral
tegmental area, nucleus accumbens, medial prefrontal and orbitofrontal cortices, hip-
pocampus, and the amygdaloid complex.
Introduction
To begin to answer the question posed in the title, we must first consider what
intrusive thinking is. Specifically, we need to address the nature of the con-
tent of an intrusive thought: Is it an episodic representation? Could it be a
visceral urge, an emotive state that is devoid of episodic content? There is
unlikely to be a single answer, as intrusive thinking is ultimately made up of
different compositions of these extremes, specific to the underlying pathology.
For example, delusions in schizophrenia are clearly episodic in nature, often
composed of a detailed and elaborate narrative. By contrast, in obsessive-com-
pulsive disorder (OCD), obsessions come more in the form of an urgency state
that is often likened to anxiety. While there may be episodic aspects to this
process, such as obsessing over specific contexts, the debilitating qualities of
12 P. E. M. Phillips and A. L. Milton
the obsession emerge from the anxiety-like state. Likewise, obsessive behavior
in substance use disorders and other addictive disorders comes in the form of
craving, an emotive state which, in its fundamental form, is devoid of episodic
content. This state, however, is intimately attached to episodic representation
of experiences related to the addictive behavior. Further, posttraumatic stress
disorder (PTSD) is also associated with an anxiety-like state that is very clearly
linked to fragmented episodic memories. This differentiation between the con-
tent of intrusive thoughts may, on the surface, seem subtle but is likely to be
critical when its neural substrates are considered because the circuits that pro-
cess episodic information are separable from those underlying emotive states.
These differences fall into a common dichotomy in the control of behavior;
namely, there are parallel systems that subserve cognitive functions. This di-
chotomy has a classic separation of processes that can be loosely summarized
in the form of a speed–accuracy trade-off: fast, low-computational processes
that amount to estimations are generated in parallel to more precise computa-
tions that have a higher cognitive demand. We and others have equated this
dichotomy to parallel processes used in machine learning that are classified as
model-free and model-based computations (Clark et al. 2012) as we believe
this is an intuitive and tractable framework. The basis of this separation is the
complexity of the information that is stored to support a learned association.
A model-based computation establishes an environmental model that can be
used to explore potential and inferred connections between stimuli and states.
In contrast, a model-free computation is a single-dimensional value assigned
to a stimulus based on the reliability of its association with a motivationally
relevant outcome.
Central to this line of reasoning, intrusive thoughts are often triggered by
environmental stimuli. For instance, in substance use disorders, drug craving
can be elicited by drug cues and is posited to become stimulus bound during the
transition to addiction (Tiffany and Carter 1998). Accordingly, drug seeking in
rodent models, a proxy for craving, can be elicited by unconditioned stimuli
(e.g., abused substances or stressors) or conditioned stimuli (e.g., drug cues or
conditioned stressors). In PTSD, intrusive episodes are often linked to envi-
ronmental stimuli in a manner consistent with overgeneralization (i.e., when
otherwise neutral stimuli elicit a threat response). Hence, intrusive thoughts
often take the form of Pavlovian-conditioned responses, which is our focus in
this chapter. As an aside, it is worth noting that while intrusive thoughts can
be triggered by unconditioned stimuli, they are unlikely to be unconditioned
responses because they are not elicited in naïve individuals but rather develop
with psychiatric pathology. Importantly, where tested, stimulus-driven intru-
sions exhibit similar neural signatures to those that are not triggered by an
explicit external stimulus (M. C. Anderson, pers. comm.).
What neural circuits are necessary to support these associations? We will
discuss circuits that support Pavlovian associations, both for emotive re-
sponses and those that incorporate representation of stimulus properties. We
Circuits That Mediate and Update Intrusive Thinking 13
will also consider structures that mediate the storage and retrieval of episodic
memories as well as the interactions between all of these circuits. We will
make the case that circuitry, including midbrain dopamine neurons that project
to the nucleus accumbens in the ventral striatum (mesolimbic pathway) or to
the medial prefrontal cortex (mesocortical pathway), is necessary for at least
a subset of emotive associations. Thereafter, discussion of the neural circuitry
will be broadened to include substrates that support Pavlovian associations
that can support inferential reasoning, with a specific focus on the central role
of the orbitofrontal cortex (OFC) in these higher cognitive processes. We will
also explore medial temporal and frontal structures implicated in the acqui-
sition and consolidation of episodic memories and discuss circuits involved
in the process of updating existing associations, both through extinction and
memory reconsolidation.
Mesocorticolimbic Dopamine- and Emotive-

Conditioned Responses
Many psychiatric disorders that are associated with intrusive thinking (e.g.,
schizophrenia, substance use, OCD, and PTSD) have been linked to perturba-
tions in dopamine transmission in the striatum and/or the prefrontal cortex.
These clues have driven extensive research on dopamine transmission with
regard to psychiatric disorders, including those where intrusive thinking is a
prominent feature.
Reward Prediction Errors
In the mid- to late 1990s, computational neuroscientists and computer scien-

tists who shared an interest in learning algorithms came to the hypothesis that
dopamine transmission provides a critical teaching signal for stimulus–re-
ward associations in a model-free learning algorithm (Barto 1995; Montague
et al. 1996). This notion was most famously linked to the empirical research
of Wolfram Schultz et al. (1997). While this area has expanded enormously
since then, the computational role of dopamine transmission in reward learn-
ing is most commonly ascribed to variants of the temporal difference algo-
rithm (Sutton and Barto 1998). This algorithm is a time-derivative model that
evolved from simple trial-by-trial learning models developed to account for
animal behavior. The hypothesized role of dopamine neurons in the model is
to signal discrepancies between reward expectation and rewards received to
update the current expectation of reward: firing increases when rewards are
larger than expected and decreases when rewards are smaller than expected.
This model puts dopamine in a critical role in the acquisition and maintenance
of stimulus–reward associations. In addition to this putative role in learning,
mesolimbic dopamine transmission has consistently been shown to invigorate
the enactment of conditioned responses (Flagel et al. 2011b; Ostlund and

Maidment 2012).
Model-Free and Model-Based Processes
In a simple cue–reward learning task, where there is spatial separation of the

cue and the reward, some animals will approach the cue during its presentation
(“sign tracking”) whereas others will approach the place where the reward will
subsequently be delivered (“goal tracking”). This latter conditioned response
requires cognitive representation of the spatial location of future reward, con-
sistent with a model-based process. Interestingly, when animals emit these
model-based conditioned responses, dopamine release in the nucleus accum-
bens does not follow the canonical prediction error signal (Flagel et al. 2011b).
This suggests that the specific pattern of signaling is selective for model-free
computations. Furthermore, while intact dopamine transmission is necessary to
perform either the conditioned response or to acquire sign-tracking conditioned
responses, it is not necessary to acquire goal-tracking responses. Similarly, dis-
rupting dopamine transmission during the Pavlovian-to-instrumental transfer
disrupts general transfer (invigoration) but it does not affect the cue’s ability
to bias action selection specifically toward the cue-associated reward, which
requires a model-based representation. Recent experiments, however, have re-
vealed a role for dopamine-encoded prediction error signals in some model-
based processes (Sharpe et al. 2017). Specifically, it was demonstrated that
phasic dopamine signals participate in the updating of the stimulus–stimulus
association that takes place in the absence of motivationally valent stimuli
and can be used to generate model-based inferences (sensory precondition-
ing). Thus, mesolimbic dopamine transmission has a somewhat nuanced role
in Pavlovian processes, with some predilection for simple stimulus–reward as-
sociations. It appears to be important in the acquisition and updating of model-
free stimulus–reward associations as well as some, but not all, model-based
associations. In addition, mesolimbic dopamine has universal psychomotor-
activating properties by which it invigorates the response to conditioned stim-
uli; however, through this, it is thought to convey only the emotive properties
of the association rather than specific (sensory) properties of the conditioned
or associated unconditioned stimulus.
Aversive Signaling
To date, the focus has been on associations with appetitive stimuli. Needless
to say, intrusive thoughts are often evoked by aversive stimuli. The role of
the mesolimbic and mesocortical dopamine systems in computations relating
to aversive information has been much more controversial. In many cases,
it is simply assumed that aversive stimuli will be computed in this learn-
ing model in a similar manner to stimuli that predict lower than previously
expected rewards. However, the evidence for this type of encoding is mixed.
Mirenowicz and Schultz (1996) observed minimal responses to mildly aversive
stimuli (air puffs to the hand) in midbrain dopamine neurons of nonhuman pri-
mates. In contrast, Matsumoto and Hikosaka (2009) observed robust changes
in the activity of dopamine cells on the presentation of aversive stimuli. In
this latter work, the investigators reported some populations of dopamine neu-
rons that encoded aversive information by changing their firing rates in the
opposite direction to reward information. Specifically, the presentation of un-
expected aversive stimuli or conditioned stimuli that increased the expectation
of aversion resulted in reduced firing of these neurons. However, Matsumoto
and Hikosaka also reported populations of putative dopamine neurons that in-
creased their firing rate to predictions of aversion. These neurons tended to
reside in dorsolateral aspects of the dopaminergic ventral midbrain (dorsolat-
eral substantia nigra pars compacta). They responded similarly to stimuli that
increased the expectation of aversion to those that increased the expectation of
reward, an encoding pattern often referred to as an unsigned prediction error.
However, this coding pattern is not without controversy. Specifically, Fiorillo
(2013) has argued that the observed positive responses to aversive stimuli re-
late to their sensory properties rather than their motivational salience. Others
have also questioned whether all of the recorded neurons in these studies are
truly dopamine-containing neurons. To address this concern of neuronal-type
specificity, Cohen et al. (2012) recorded the firing rates of genetically identi-
fied dopamine neurons in mice in response to presentations of appetitive- and
aversive-related stimuli. They reported that modulation of the firing rates to
aversive-related stimuli were consistently in the opposite direction to those
for reward stimuli. These neurons, however, were exclusively recorded in
the ventral tegmental area (i.e., the ventral medial aspect of the ventral mid-
brain) and so did not include the homologous anatomical region from which
the unsigned prediction error signals were reported by the Hikosaka group.
In neurochemical studies that measure dopamine levels in terminals, results
with aversive stimuli have also been mixed. Studies measuring dopamine lev-
els over minutes tend to report increases in dopamine to the presentation of
aversive stimuli, especially in the prefrontal cortex (Young 2004; Butts et al.
2011). In contrast, measurements on the order of seconds reveal decreases in
dopamine in the nucleus accumbens to aversive stimuli (Roitman et al. 2008).
Some investigators addressing this issue have argued that increases in dopa-
mine transmission following an aversive stimulus observed over minutes were
responses to the relief from aversion at the offset of the stimulus rather than a
response to the onset (Ungless 2004).
Integration of Appetitive and Aversive Information
With systematic interrogation of dopamine neurons in the ventral tegmental

area based on identified afferent and efferent connectivity, Lammel et al. (2012)
proposed that appetitive and aversive processing by dopamine neurons is seg-

regated into subcircuits. They demonstrated that activation of a circuit con-
necting the lateral habenula, ventral tegmental area, and the medial prefrontal
cortex produces an aversive state, while activation of another circuit connect-
ing the laterodorsal tegmentum, ventral tegmental area, and lateral shell of
the nucleus accumbens produces an appetitive state. Complementary to these
data, Tye and colleagues recently reported that a set of prefrontal-projecting
dopamine neurons were selectively activated by aversive stimuli, whereas do-
pamine neurons that projected to the nucleus accumbens were activated by
rewards (Vander Weele et al. 2018). While this schema does not account for all
of the apparent discrepancies in the literature, it does move the field forward
toward a resolution. However, the issue of whether and how dopamine trans-
mission integrates appetitive and aversive information in a manner that could
instantiate a single, unitary learning model is still an open question. Given
our current understanding, this integration could be through one of several
possibilities, including bidirectionality of appetitive and aversive information
by mesolimbic dopamine neurons (Roitman et al. 2008; Cohen et al. 2012),
gleaning appetitive and aversive information from different populations of do-
pamine neurons (Lammel et al. 2012; Vander Weele et al. 2018), or combining
appetitive information from dopamine neurons with aversive information from
other neural substrates (Daw et al. 2002).
Broader Circuity
When considering the entirety of the pathways which support these stimu-
lus–stimulus associations, the complexity of the circuitry rapidly expands. In
addition to the laterodorsal tegmentum and the lateral habenula, the central
nucleus of the amygdala, in particular, has been implicated as important up-
stream circuitry to dopamine neurons in model-free processes (Clark et al.
2012). Nonetheless, it is important not to dismiss the rich convergent inputs
into the ventral midbrain from many areas of the brain (Geisler and Wise
2008). Indeed, an optimal temporal difference algorithm should have access to
all available sources of predictive information about rewards and punishers as
well as information on motivational states.
More Complex Stimulus–Stimulus Associations
Orbitofrontal Cortex
From a plethora of research, it is evident that the OFC encodes many differ-
ent features of motivational stimuli (Thorpe et al. 1983; Padoa-Schioppa and
Assad 2006; Stalnaker et al. 2014). Indeed, it seems that just about any as-
pect of perception is encoded in about twenty percent of OFC neurons! This
multidimensional encoding seems to be especially important to support model-

based associations that permit inferential reasoning. For example, OFC lesions
do not affect the acquisition of simple Pavlovian associations or the ability
for animals to avoid food that has been paired with illness or fed to satiety.
However, unlike intact controls, OFC-lesioned animals continue to respond
to stimuli that predict these devalued outcomes (Gallagher et al. 1999). These
reinforcer devaluation studies demonstrate that the OFC contributes to the ani-
mal’s ability to derive the updated expected value of a cue by linking the previ-
ously learned stimulus–reward association with the current incentive value of
that outcome without having yet directly experienced the pairing of this cue
with the devalued outcome. Consistent with the OFC playing a selective role
for model-based computations, OFC lesions do not affect general Pavlovian-
to-instrumental transfer, but they do disrupt the cue’s ability to selectively
enhance responding for the specific outcomes the cues predict (Ostlund and
Balleine 2007a). While the computational role of the OFC in these processes is
not yet precisely defined, a reasonable inference would be that the OFC is used
to evaluate the content of model-based associations upon their retrieval rather
than being directly involved in the acquisition or storage of this information
per se. In this regard the OFC and dopamine may have parallel, complemen-
tary roles in response to conditioned stimuli with dopamine conveying emotive
responses and OFC relaying stimulus-specific sensory information.
Episodic Information
While the above discussion differentiates associations that represent specific

features of conditioned or unconditioned stimuli, these more complex asso-
ciations do not necessarily incorporate episodic information. However, since
intrusive thoughts are often episodic in nature, it is pertinent to expand our
consideration to circuits involved in the processing of episodic information.
The study of episodic memory is a particularly rich area of neuroscience and
has primarily focused on the medial temporal lobe, including hippocampus
formation as well as the prefrontal cortex and to some extent the frontal and pa-
rietal cortices. However, investigations of the interactions between Pavlovian
associations and episodic memory is relatively sparse. Nonetheless, a particu-
larly elegant series of experiments from Shohamy and colleagues have pro-
vided evidence that the hippocampus has central roles in associative processes.
For example, they showed that hippocampal episodic information can pro-
vide a framework for model-based processes to permit inferential reasoning
(Wimmer and Shohamy 2012). They also demonstrated that the hippocampus
can drive dynamic corticostriatal network connectivity governing stimulus–re-
ward associations (Gerraty et al. 2018). These innovative studies have started
to build a foundation for understanding the use and integration of episodic in-
formation into learned associations. This platform could be particularly useful
for studies of intrusive thinking.
Updating and Inhibition of Intrusive Thought
Extinction
To think about the updating of intrusive thoughts, we should consider the dif-
ferent plasticity processes that can be engaged following the consolidation of
an intrusive thought or memory. Broadly speaking, the neural trace or “en-
semble” that encodes a thought or memory can be triggered to induce retrieval,
reconsolidation, or extinction of the trace under similar, but importantly differ-
ent, conditions. Extinction has been the most extensively studied, having been
originally described by Pavlov (1927). Extinction is operationally defined as
the degradation of behavior that was previously supported by a learned asso-
ciation. It takes place when the reliability of the association is weakened, typi-
cally occurring after extensive exposure to the unreinforced cue. Extinction
can be modeled as a reduction in the strength of an association between a con-
dition and unconditioned stimulus in a simple bidirectional learning system,
such as the model-free system putatively associated with dopamine transmis-
sion. However, for many associations, it has been argued that extinction learn-
ing is not simply the “unlearning” of an association but new, discriminative
learning that associates different states of the world with new contingencies in
a more complex model. While there are clear demonstrations that extinction
can be new learning—most likely dependent on prefrontal cortical regions—it
perhaps should not be assumed that this is a universal mechanism. At least
at the level of the amygdala, molecules usually associated with the depoten-
tiation of synapses increase in activity during Pavlovian extinction, and these
molecules are also necessary for extinction to occur (Merlo et al. 2014). Thus,
it remains a possibility that different associations have fundamentally differ-
ent mechanisms of extinction. Regardless of mechanism, a defining feature of
intrusive thinking is that the underlying associations are relatively resistant to
extinction.
Reconsolidation
The resistance to extinction of intrusive memories may be especially maladap-

tive if, instead of engaging extinction, reactivation of the intrusive thought
leads instead to strengthening of ensemble via reconsolidation mechanisms.
Reconsolidation is a process, more recently described than extinction, that can
also potentially act to update memories. Reconsolidation refers to the induc-
tion of memory lability (or, more mechanistically, memory “destabilization”)
under certain conditions of retrieval, and the subsequent “restabilization” of
the trace in a strengthened or updated form. The restabilization phase is depen-
dent upon protein synthesis in a similar manner to the original consolidation
process when the memory was first stored. Therefore, preventing reconsolida-
tion following memory retrieval (e.g., by administration of a protein synthesis
inhibitor) can essentially erase the memory. Reconsolidation has been investi-
gated in the context of fear memories (Nader et al. 2000), where it was shown
to be dependent on protein synthesis in the basolateral amygdala. Likewise,
the basolateral amygdala is necessary for the reconsolidation of drug-related
associations that support place conditioning and conditioned reinforcement
(Milton et al. 2008a, b; Théberge et al. 2010). Interestingly, protein synthesis
in the nucleus accumbens core is also required for reconsolidation of asso-
ciations supporting conditioned drug place preference but is not required to
reconsolidate drug-cue associations that support conditioned reinforcement
(Théberge et al. 2010). The extent to which the disruption of reconsolidation in
one node of a motivational network can impinge on the function of other nodes
in that network is a question that has been surprisingly understudied. Despite
this, it has been hypothesized that reconsolidation could provide a mechanism
by which memories can be strengthened (Lee 2008), generalized (Vanvossen
et al. 2017), and integrated into wider memory networks (Hardt et al. 2010).
Therefore, if reconsolidation mechanisms were engaged upon reactivation of
an intrusive thought, it is possible that a form of mnemonic “positive feed-
back” could be established, by which a reactivated intrusive thought not only
becomes strengthened when it restabilizes, but generalizes and integrates with
other associative traces to lead to an increase in the number of cues and con-
texts that could trigger the intrusive thought. The consequent increase in the
likelihood of triggering the thought would potentially increase the likelihood
of subsequent reactivation, leading to further strengthening, generalization, in-
tegration, and so on.
Therapeutic Strategies Using Updating Mechanisms
In addition to providing a potential pathological mechanism underlying the

persistent and recurrent nature of intrusive thought, reconsolidation could
also provide a therapeutic target. While the administration of protein synthe-
sis inhibitors to humans to disrupt reconsolidation is not straightforward, it is
possible to capitalize on the updating function of memory reconsolidation in
designing therapeutic approaches. For example, developing an approach first
used in the preclinical literature (Monfils et al. 2009), Schiller et al. (2010)
used a “retrieval-extinction” procedure in which they reactivated a fear mem-
ory and subsequently extinguished this memory while it was destabilized (in
the “reconsolidation window”). Similarly, James et al. (2015) used visuospatial
interference (playing the video game Tetris) to disrupt the reconsolidation of
intrusive mental images produced by watching traumatic films clips. Although
further research needs to be conducted to determine the mechanisms by which
these procedures produce effects, including corroboration at the molecular
level (Cahill and Milton 2019), these approaches hold potential for the de-
velopment of new treatment for neuropsychiatric conditions characterized by
intrusive thoughts. As a cautionary note, an important consideration for these
types of approaches is that different types of stimulus–stimulus associations

(i.e., emotive and cognitive) could exist in parallel, supported by independent
neural circuits. Therefore, which memory or, more specifically, which aspect
of the association drives the pathology may be critical. With some pathologies,
for instance, extinguishing an intrusive episodic memory could be futile if the
untreated emotive association reattaches to a new episodic memory.
Conclusions
The neural circuits that mediate and update intrusive thoughts are complex but
potentially tractable based on our current understanding of model-based and
model-free systems and their operation. It is important to appreciate, however,
that these circuits are not fixed and immutable, but rather it is likely that they
undergo repeated rounds of plasticity and metaplasticity, leading to imbalance
within the circuit. In this way, we hypothesize that an adaptive physiological
process, supported by functional neural circuitry, can become persistent, recur-
rent, and pathological.
Acknowledgments
This work was supported by NIH grants R01-DA039687, U01-AA024599, and P50-
MH106428-5877 to Paul Phillips, and U.K. Medical Research Council Programme
Grant MR/N02530X/1 to Amy Milton. Amy Milton is further supported by the Ferre-
ras-Willetts Fellowship in Neuroscience at Downing College, Cambridge.
3
Corticostriatal Intrusions
Bernard W. Balleine
Abstract
The loop-like circuits which link the cortex and basal ganglia have been implicated in
a range of functions; most recently in the precursors to movement, including planning
and decision making. Damage to these circuits induced by various disease states have,
therefore, been heavily implicated in a range of symptoms, including intrusive involun-
tary thoughts and actions associated with, for example, neurodegenerative and psychi-
atric conditions as well as addictions of various kinds. This chapter focuses on recent
evidence of parallel circuits that mediate the distinct forms of control associated with
reflexive and volitional actions, and the interactions between these circuits in determin-
ing adaptive behavior. It discusses two kinds of interaction important for understanding
intrusive actions and thoughts: competitive interactions, whereby circuits controlling
volitional actions regulate reflexive or habitual responses, and cooperative processes
that allow the simulation of specific actions to become manifest in performance. It then
explores the role of information derived from predictive learning in action selection and
choice. The influence of such information is conveyed through a specific corticobasal
ganglia circuit, damage to which has been implicated in compulsive action. The evi-
dence considered generally suggests that intrusive thoughts and actions are the product
of an imbalance between corticobasal ganglia circuits rather than dysfunction in any
one circuit or its related control process.
Introduction
Intrusive thoughts and actions are highly debilitating symptoms associated

with a range of psychiatric disorders and addictions. There are numerous
theories regarding the psychological, behavioral, and neural determinants of
such intrusions; however, in recent years, a growing theme has been to link
these conditions to the processes associated with involuntary or reflexive ac-
tions, commonly known as habits (Everitt and Robbins 2016; Robbins et al.
2019). The distinction between volitional, goal-directed actions and reflexes
is one with which students of adaptive behavior have grappled for millen-
nia. If, today, this issue feels more tractable, it is because of advances in
our understanding of the behavioral processes and, consequently, the neural
22 B. W. Balleine
networks that support these forms of action control. For the purposes of this
chapter, I will focus on two aspects of these networks that appear to have
important implications for future research into these issues. The first lies in
the relationship between the neural networks that mediate goal-directed and
habitual control: how they compete and cooperate to support our everyday
activities, and the consequences of failures in these interactive processes.
The second lies in the way an even more fundamental reflexive system me-
diating Pavlovian-conditioned reflexes interacts with goal-directed processes
to provoke the initiation and performance of otherwise volitional actions in
ways that can often appear maladaptive. Although both undoubtedly relate
to the habitual control of intrusions, neither wholly depends on such factors.
Both, however, highlight how actions and habits are integrated and, ulti-
mately appear to suggest that failures, when they occur, are failures in that
integrative process.
Actions and Habits
It is now well recognized that the performance of instrumental actions (i.e.,

those actions through which we manipulate the environment) is subject to two
forms of control process, typically referred to as goal-directed and habitual
control. Goal-directed actions are determined by their relationship to and the
value of their consequences or outcome. They are rapidly acquired through a
learning process sensitive to the causal relationship between action and out-
come, and they uniquely engage a process of cognitive-emotional integration
in linking these causal relations to goal values based on the motivational and
emotional processes through which the incentive values of specific outcomes
are established (Dickinson 1994; Dickinson and Balleine 1994; Balleine 2001).
As such, goal-directed control is effortful and costly; it engages considerable
cognitive and emotional resources but provides fast solutions that are highly
flexible (i.e., they can be executed or withheld on demand). Habits, on the
other hand, are determined by their antecedents rather than their consequences.
They are acquired more slowly through a process of sensorimotor association
with specific associations selected by reinforcement. They require few, if any,
cognitive resources for their acquisition or performance and are emotionally
subject to a reinforcement signal (during acquisition) and to the net level of
arousal or drive (during performance) (Dickinson and Balleine 2002). Habits
are released, rather than executed, by environmental events and are difficult to
withhold once released. They are, however, highly organized actions that have
a similar topography across repetitions; thus, they make an adaptive, low-cost
solution to common or routine problems. Importantly, these two forms of ac-
tion control have not only been found to be mediated by distinct learning rules,
associative structures, and emotional feedback processes (cf. Dickinson 1994);
they also engage distinct neural processes.
Corticostriatal Intrusions 23
Common Tests of Goal-Directed Action
Much of the evidence for distinct forms of action control has been derived
from the use of a battery of tests of goal-directed control through which the
sensitivity of animals, whether rodent or human, to changes in the action–
outcome relationship and in outcome value is evaluated (see Dickinson and
Balleine 1994; Balleine and Dickinson 1998; Balleine 2005). One such test,
the contingency degradation test, assesses sensitivity to the relationship be-
tween an action and its consequences by increasing the probability of access,
p, to a specific outcome, O, in the absence of an action, A, or p(O|noA). Given
a specific probability that an action earns an outcome (i.e., p(O|A)), p(O|noA)
can be increased until the outcome is equally probable if the action is per-
formed or not. At that point, the action is no longer causally related to the
outcome and, if the actor is sensitive to this, then the actor should no longer
perform the action (Hammond 1980). Increasing p(O|noA) beyond that point
means that the action prevents the outcome, which should reduce responding
even faster, something demonstrated through the use of omission schedules
(Davis and Bitterman 1971). Goal-directed action in humans and animals is
exquisitely sensitive to these manipulations of the action–outcome relation-
ship, largely because these manipulations alter the causal relationship between
these events (Dickinson 2012). As a consequence, not only is performance
affected, judgments as to how causal an action is, with respect to its specific
outcome, are similarly modified by these contingency manipulations (Shanks
and Dickinson 1991). Finally, these changes in contingency are highly selec-
tive; altering the relationship between one action and outcome does not affect
performance based on the causal relationship between other actions and their
outcomes (Balleine and Dickinson 1998).
The second commonly used test is called outcome devaluation, which as-
sesses sensitivity of performance to changes in the value of the outcome of an
action and is usually conducted without feedback (i.e., in extinction). Hence,
having learned a number of specific action–outcome relationships (A1 → O1,
A2 → O2, etc.), the value of one outcome can be altered, after which the sub-
ject is given a choice between the various actions. Goal-directed control re-
flects the ability to integrate the action–outcome relationship with the altered
value of the outcome to modify performance of the action. Again, considerable
evidence has demonstrated that goal-directed actions in both humans and ro-
dents are sensitive to these kinds of treatment (e.g., Balleine and Dickinson
1998; Balleine and O’Doherty 2010).
Neural Bases
Studies of goal-directed action have found evidence that the acquisition and
performance of such actions depend on the rich connections of the prefrontal
cortex (PFC), including the human ventromedial PFC and medial orbitofrontal
24 B. W. Balleine
cortices, with the striatum, specifically the caudate nucleus (Balleine and
O’Doherty 2010). The greater resolution allowed by rodent studies has signifi-
cantly refined this general picture. Thus, it is now clear that the homologous
region of the medial PFC in rats, the prelimbic cortex (area 32), is engaged
during the early acquisition of goal-directed actions (Hart and Balleine 2016).
Furthermore, whereas the input layers 2/3 contribute significantly to the con-
solidation of goal-directed learning, the output layers (particularly the intratel-
encephalic neurons in layer 5) are critical to this process. Activity in their
bilateral projection is necessary for learning-related plasticity in the posterior
dorsomedial striatum (one of their main targets) and thus for the acquisition of
goal-directed actions (Hart et al. 2018a, b).
Using the above tests, we demonstrated, some time ago, that rodents with
damage to the prelimbic cortex, the posterior dorsomedial, or the mediodorsal
thalamus show deficits in their sensitivity to contingency degradation and out-
come devaluation, and continued to respond as if neither the contingency nor
outcome value had changed. Evidence that these structures share a high degree
of interconnectivity led to the claim that they constitute the critical cortico-
striatal-thalamo-cortical loop through which goal-directed actions are encoded
(see Figure 3.1a; Balleine 2005). Importantly, this same insensitivity to shifts
in the action–outcome relationship and in outcome value is found in habitual
actions. Early in training, actions are sensitive to changes in contingency and
value, whereas with continued practice, performance naturally becomes in-
sensitive to these changes as action control shifts from the consequences of
an action to antecedent environmental stimuli with which the action becomes
associated (Dickinson et al. 1995, 1998). Chief among these stimuli is the con-
text in which the action is trained (Thrailkill and Bouton 2015).
At a neural level, the acquisition and performance of habits depends on a
corticostriatal circuit linking the sensorimotor regions of frontal cortex with
the putamen or dorsolateral striatum (Figure 3.1a). Again, treatments causing
degeneration to, or temporary inactivation of, structures in this circuit block
the acquisition of habit learning and render habitual actions goal directed; that
is, despite extensive training, they continue to show sensitivity both to shifts
in the action–outcome relationship and outcome value (Yin et al. 2004, 2006).
Generally, therefore, these findings have been interpreted as suggesting that
the goal-directed and habitual control of instrumental actions is a competi-
tive process: any reduction in goal-directed control increases the likelihood ac-
tions will be habitual, whereas any reduction in habitual control increases the
likelihood actions will be goal directed. Thus, inactivation of the dorsomedial
striatum immediately places actions under habitual control (Yin et al. 2005a),
whereas inactivation of the dorsolateral striatum appears immediately to ren-
der actions goal directed (Yin et al. 2006), as if these two processes are always
active and merely compete for control over performance (Balleine et al. 2009).
In fact, even under normal circumstances the goal-directed process can rap-
idly inhibit habitual control. This can be detected in our everyday activities. A
(a) Cognitive control Stimulus control
(c)
Goal-directed action Habitual action
S – R Xecute
MPC cortex SM
Senstive to changes Insenstive to changes
in their causal in their causal select
consequences consequences
Sensitive to changes Insensitive to changes
in goal value DMS striatum DLS in goal value
A – Ov
Rapidly acquired Slowly acquired
Flexibly deployed Inflexibly deployed evaluate
MD thalamus PO
Necessary for Impulsive and
inhibitory control perseverative
Habit memory Associative memory
SN midbrain GPi (d) …
S1 S2 Sn A S
(b)
R1 R2 … Rn
GDA Habit O
+ –
Arbitrator
R1 R2 Rn Evaluative memory
Motor output
Figure 3.1 Competition and cooperation between actions and habits. (a) Schematic of neural circuits mediating goal-directed and habitual ac-
tion control, as well as their primary characteristics: A loss of goal-directed control tends to yield dysregulated habitual actions, whereas a loss of
habitual control yields dysregulated goal-directed actions. (b) One view of the competitive processes in goal-directed action (GDA) and habitual
action control in which both compete for control of performance via some form of arbitration; habits emerge via arbitration but are heavily regu-
lated. (c) Illustration of the cooperation between the stimulus-mediated selection and action-mediated evaluation processes necessary to generate
action execution. Generally, stimuli (S) generate urges to respond (R) that initiate the retrieval of specific actions (A), their outcomes (O), and the
value (v) of those consequences. If negative, this evaluation would check that urge; if positive, the urge would translate into an executed action. (d)
This descriptive model can be further elaborated into an associative cybernetic model which views instrumental performance as the direct outcome
25
of cooperation between S-R and A-O associative processes, whose joint influence sums at the motor system to drive motor output.
26 B. W. Balleine
useful example is to consider the effect of seeing a police car in the rearview
mirror while driving on the freeway: After a period of carefree, apparently
cognitively disconnected driving, the sight of the highway patrol causes a
very distinct change in our behavior. Do we carry on driving in such a care-
free manner? Not likely! Even if we are within the speed limit and are gener-
ally obeying the rules of the road, our vigilance would dramatically increase
and our driving would become far more deliberate. Thus, habitual control has
been suppressed. Likewise, rats behaving habitually will rapidly transition
to goal-directed control when the lever response is punished by the actual
delivery of an aversive or noxious outcome (Dickinson et al. 1983, 1995).
The rapidity of this adjustment is, however, severely curtailed by damage to
or inactivation of the dorsomedial striatum (Yin et al. 2005b; Furlong et al.
2017), a finding that is consistent with the argument that the return of control
to the goal-directed system is the source of the rapid suppression of habits in
a punishment situation (see Figure 3.1b).
Loss of Control and Intrusive Action
Some forms of psychopathology that result in intrusive or compulsive actions

find their source in a defective ability to suppress habitual actions; most no-
tably in drug addiction. During the development of addiction, the pursuit of
drugs of abuse rapidly becomes habitual, coming under the control of internal
and external states and stimuli rather than the consequences of acting (Corbit
et al. 2014; Furlong et al. 2016, 2017, 2018). It is important, however, to dis-
tinguish habitual drug seeking from other forms of habitual behavior. Under
normal conditions, habit learning can be highly adaptive: habits allow us and
other animals to relegate the control of routine behavioral responses to a sys-
tem that uses few cognitive resources, freeing up a limited executive capacity.
In contrast, habitual drug seeking is pathological: drug exposure increases the
rate of acquisition of habitual actions and the influence of drug-associated con-
texts (Ostlund et al. 2010) and cues (Glasner et al. 2005) on their performance.
Furthermore, a distinguishing feature of habitual drug seeking is the addict’s
loss of executive control over the habit. As is commonly noted, drug pursuit
persists in the face of often severe, negative consequences. The compulsive
pursuit of drugs can be viewed, therefore, as the product of (a) a drug-induced
increment in habit and (b) a decrement in the addict’s ability to exert con-
trol over these actions in the face of persistent, negative feedback (Ostlund
and Balleine 2008). The need for these two processes to interact may help
clarify why relatively few people who take drugs ultimately become addicted;
although taking drugs may have some of the hallmarks of a habit, resilience in
their goal-directed system to the effects of drug exposure may help to reduce
the risk of those habits ultimately becoming dysregulated (Burt et al. 2016).
Consistent with this argument, we have found that exposure either to drugs
themselves, such as cocaine or methamphetamine, or to contexts previously
paired with drugs can render goal-directed action habitual in tests of outcome
devaluation (Ostlund et al. 2010; Corbit et al. 2014; Furlong et al. 2017).
Furthermore, actions in these contexts can persist when punished by the de-
livery of an aversive outcome (Furlong et al. 2018), something that we found
was linked to reduced activity in D1-expressing spiny projection neurons
(SPNs) in the posterior dorsomedial striatum (Furlong et al. 2017). There are
two populations of SPNs in the striatum that account for roughly 95% of the
neurons: direct path D1-expressing (dSPNs) and indirect path D2-expressing
SPNs (iSPNs). Whereas dSPNs tend to increase functional output, and hence
increase goal-directed control, iSPNs inhibit it (Gerfen and Surmeier 2011).
As such, a loss of dSPN activity should be expected to have the effects we
observed. In an attempt to rescue normal function, therefore, we attempted to
redress the relative balance between the activity of dSPNs and iSPNs by in-
hibiting iSPNs. How we achieved this was complex. The binding of dopamine
at D2 receptors on iSPNS inhibits the activity of these neurons and, as such, a
D2 antagonist might be expected to be sufficient. However, D2 receptors are
also expressed on multiple interneurons as well as iSPNs in the dorsal striatum,
reducing the selectivity of this manipulation. Importantly, however, adenosine
A2A receptors are only expressed on iSPNs in this region, and the inhibition
of these receptors increases the effects of D2 binding only on iSPNs (Tozzi et
al. 2007). We hypothesized, therefore, that local infusion of an A2A antagonist
would (a) increase D2 receptor activity, (b) reduce iSPN activity, (c) restore
the balance between dSPNs and iSPNs, and thus (d) allow the animals to exert
behavioral control over their instrumental performance. This is indeed what we
found: rats were now able to exert behavioral control over their habits and re-
duced performance in a punishment test to a similar degree whether they were
tested in a drug-paired or unpaired context (Furlong et al. 2017).
Similar deficits in contingency degradation and outcome devaluation have
been described in various psychiatric conditions linked to changes in the cir-
cuitry mediating goal-directed action control (Griffiths et al. 2014). We found,
for example, that the causal judgements of a cohort of youths diagnosed with
major depression were relatively insensitive to changes in the causal relation-
ship between action and outcome (Griffiths et al. 2015, 2016a). Furthermore,
reductions in causal awareness were correlated with size and shape changes of
the globus pallidus (GP) and midline thalamic structures in the basal ganglia
output circuit that feeds back to the cortex. Tractography confirmed the rela-
tionship between the dysfunctional area of GP and the striatum, on one hand,
and the midline thalamus, on the other (Griffiths et al. 2015). This suggests,
particularly given that subjects were at an early stage in illness progression,
that such changes may reflect the precursors of later prefrontal cortical and
corticostriatal deficits in depression, as has also been claimed by others with
regard to schizophrenia (e.g., Simpson et al. 2010). Indeed, in a recent study,
we found deficits in the flexibility of causal judgment in chronic schizophrenia
similar to that observed in major depression (Morris et al. 2018). Furthermore,
28 B. W. Balleine
we found that schizophrenic subjects had a severe deficit in the sensitivity

of their instrumental performance to outcome devaluation. We did not, how-
ever, find deficits in the ability of schizophrenic subjects to describe the im-
mediate relationship between actions and their consequences, nor did we find
that schizophrenic subjects differed from healthy controls in their ratings of
outcome value after devaluation. Rather, it appeared that the schizophrenic
subjects had difficulty integrating their knowledge of the action–outcome re-
lationship with the changes in outcome value to permit them to make accurate
choices. As might be expected, this effect was related to reduced activity in
the dorsolateral prefrontal cortex and in the head of the caudate. Furthermore,
whereas the deficits in neural activity were related to the severity of negative
symptoms, particularly avolition and alogia, devaluation sensitivity correlated
with functional measures, such as hours of paid employment over the last six
weeks (Morris et al. 2015). Generally, therefore, with respect to its role in
controlling intrusive actions, each of these effects suggests that dysfunction in
the prefrontal-caudate-globus pallidus-thalamic feedback circuit compromises
goal-directed control.
The Dysregulation of Goal-Directed Action
It would be interesting to know whether the same kind of dysregulation oc-

curs when habitual control is inhibited. Does reduced habitual control result in
dysregulated goal-directed action? If so, what would that dysregulation look
like? There are, in fact, layers of complexity in attempting to understand the in-
teraction of these apparently distinct action controllers. At one level it is clear
that goal-directed and habit processes compete for control of performance
(as reviewed above). There is also evidence that these processes cooperate
in the integration of stimulus-mediated action selection with action evalua-
tion, particularly during the performance of goal-directed actions. Perhaps the
strongest evidence comes from studies of instrumental reinstatement: after a
period of extinction, we have found that free delivery of the instrumental out-
come reinstates performance of its associated action (Ostlund and Balleine
2007b). Importantly, the ability of the outcome to generate this effect does not
depend on its value: devaluation of the freely paired outcome does not affect
its ability to select its associated action in reinstatement tests. Nevertheless, the
subsequent level of performance of that action is affected by the devaluation
treatment: devaluing the outcome that serves as a goal for the retrieved action
reduces the vigor of its performance but not the ability of the outcome to rein-
state performance (Balleine and Ostlund 2007; de Wit et al. 2009).
Thus, in the ordinary course of events, this evidence suggests that the outcome
controls actions in two ways: (a) through a form of stimulus–response associa-
tion, according to which the stimulus properties of the outcome select its associ-
ated action, and (b) through the action–outcome association, through which the
action retrieves the outcome as a goal. This behavioral evidence suggests that a
selection–evaluation–execution sequence lies at the heart of goal-directed instru-

mental performance and that this control requires the cooperative integration of
the goal-directed and habitual control processes (see Figure 3.1c; Balleine and
Ostlund 2007; Balleine and O’Doherty 2010). From this perspective, two pos-
sibilities may arise. First, if maladaptive habitual control results in attenuated
cooperation between the habitual and goal-directed control processes, then this
would break the link between the urge to respond and the regulation of the effect
of that urge on performance, potentially resulting in an urge to act but without
the subsequent capacity to evaluate and, when necessary, to suppress that ac-
tion. Such intrusive urges to act may be one source of obsessions and, if strong
enough, of compulsions (as discussed above). Alternatively, an actual loss of
habitual control could result in a loss of action selection and/or initiation leaving
the actor cognitively able to retrieve specific action outcomes and their value but
unable to implement those actions. Cognitively simulating actions and thinking
through their consequences sounds like planning. However, repeatedly engaging
in cognitive simulation because of an attenuated capacity to complete the re-
trieval–evaluation–execution sequence does not resemble adaptive planning but
rather obsessive or intrusive thinking. It might, therefore, be argued that whereas
attenuated goal-directed control results in compulsive or intrusive actions, a loss
of habitual control results in obsessive or intrusive rumination on the potential
consequences of actions that are, then, unable (or less able) to be performed.
Such failures could result from dysregulation in the competition between these
control processes (resulting in a loss in the ability to inhibit habits) or in the co-
operation between them (resulting in an inability to execute actions and so bring
planning to an end).
Some years ago, we implemented this general scheme within an associa-
tive cybernetic model of instrumental performance (see Figure 3.1d). Briefly, a
response tendency or urge is generated in a stimulus-response memory which
then brings to mind the retrieval of a specific action and its consequences in an
associative memory. Outcome retrieval in the associative memory activates an
incentive memory of that outcome which, by marshalling specific motivational
and emotional processes, determines its value, which acts to potentiate (or de-
potentiate) the motor effects of the stimulus-response urge and thus increase
(or decrease) the probability that the action will be executed. It is this latter
process that constitutes the cybernetic or feedback component of the model
without which either response urges could run off unchecked or not run off at
all (depending on their strength). Likewise, a significantly attenuated stimu-
lus-response memory would result in the failure of activity in the associative
memory to result in an actual response. One question that arises in this model
is why there is any activity at all in the associative memory when the habit
memory is attenuated. To understand this issue, it is necessary to introduce per-
haps the single most important influence on the performance of goal-directed
and habitual actions in the context of intrusive thoughts and actions: the effect
of predictive stimuli on action selection.
30 B. W. Balleine
Cognitive Control of Action
Although both prediction and control are necessary for adaptive behavior, dis-
cussions of the importance of predictive learning are usually confined to its
role in determining conditioned reflexes of one kind or another. There are,
however, multiple ways in which predictive learning can influence instru-
mental actions. Pavlovian predictors of important environmental events could
elicit conditioned reflexes congruent or incongruent with our actions; for ex-
ample, a stimulus paired with an aversive event could generate freezing, which
would be incongruent with active avoidance responses. Furthermore, there is
evidence that predictive learning can influence actions independently of the
conditioned reflexes predictors produce. Indeed, there is good evidence that a
stimulus that has become a reliable predictor of a valued outcome can enhance
the performance of actions associated with that outcome while leaving actions
associated with other rewarding outcomes unaffected. Furthermore, treatments
that abolish the predictive validity of such stimuli can abolish these effects on
instrumental performance without affecting their ability to evoke a conditioned
reflex (Delamater 1995). Over and above conditioned reflexes, therefore, pre-
dictive learning clearly provides information regarding forthcoming rewards
and punishers; if that information is important for adaptive behavior, its effects
will then be mediated by changes in instrumental action, via its effects on ac-
tions sufficiently flexible to be modified in response to that information and to
do so rapidly.
The influence of the information provided by predictive learning on in-
strumental action is typically studied in the laboratory using the Pavlovian-
instrumental transfer paradigm (for a recent review, see Cartoni et al. 2016). In
these experiments, subjects are exposed to a period of Pavlovian conditioning,
during which cues of various kinds are paired with specific, usually rewarding,
events (e.g., specific foods or fluids), after which they are trained to perform
distinct actions to earn those same food or fluid outcomes (Figure 3.2a). In a
typical rodent experiment, rats will be first given a period of predictive learn-
ing, during which they learn that stimuli S1 and S2 (e.g., tones or clickers)
predict the delivery of distinct outcomes, O1 and O2 (e.g., dry food pellets or
liquid sucrose): S1–O1 and S2–O2. Subsequently, they are trained to perform
two novel instrumental actions (A1 and A2) for these same outcomes. They
might be trained, for instance, to press one lever for the pellets and a second
lever for the sucrose: A1 → O1 and A2 → O2. In a final test, the previous
Pavlovian and instrumental phases are brought together to examine the influ-
ence of the former on the latter: rats must choose between A1 and A2 in the
presence of S1 and S2 (i.e., S1: R1 vs. R2, and S2: R1 vs. R2). Importantly, no
outcomes are delivered in this test phase, either after the stimuli or the actions.
Thus, the test provides an opportunity to observe how predictive learning influ-
ences instrumental performance directly. Typically, the stimuli cause the rats
to select and perform more vigorously the response previously associated with
(a) Predictive learning (c)
(Pavlovian conditioning)
mdT
Goal-directed learning
(instrumental conditioning)
Choice
(b) (d)
evaluates
selects Rn On
Medial
thalamic
Candate nuclei
Os v
VP
earns OFC
Rn On S
initiates
Figure 3.2 The influence of predictive learning on choice. (a) Design of a typical experiment used to study the influence of predictive learning
(Pavlovian conditioning) on instrumental choice performance, here in rodents. Rats are first exposed to two stimuli paired with distinct outcomes
and then trained to perform two actions for those outcomes before the influence of predictive learning on choice is assessed in an extinction test.
(b) Evidence (see text) suggests that the influence of predictive learning is mediated by the ability of a stimulus, S, to retrieve a specific outcome,
O (S – On), and so the specific action associated with that O (On – Rn). The outcome delivered as a consequence of Rn serves both as the goal of the
action and as a stimulus that selects the next action. (c) Schematic of the cortical-striatal-pallidal-thalamo-cortical circuits underlying goal-directed
action and the way predictive learning influences these actions. Note that the influence of stimuli on infralimbic and orbitofrontal cortex (IL/OFC)
activity is driven by the ability of nucleus accumbens shell (NAsh) and medial ventral pallidum (VPm) to inhibit mediodorsal thalamus (mdT),
and so disinhibit this input. The involvement of medial OFC and its connections with nucleus accumbens core (NAco) is crucial to the integration
of this circuit with the larger goal-directed circuit and its influence on performance through medial agranular cortex (M2). (d) Dysfunction in the
same cortical-striatal-pallidal-thalamo-cortical circuit in humans has been argued to underlie the compulsive actions associated with obsessive-
31
compulsive disorder; ventral tegmental area (VTA), caudate nucleus (CAUD) (after Modell et al. 1989).
32 B. W. Balleine
the outcome predicted by the stimulus: given the scenarios described above,
S1: R1 > R2 and S2: R1 < R2. Very similar effects are observed in mouse and
human subjects to those observed in rats (Cartoni et al. 2016).
As intimated above, the influence of predictive stimuli on choice in the
transfer paradigm depends on their predictive status; degrading the contin-
gency between a stimulus and its specific outcome does not affect its ability
to elicit a conditioned response but completely abolishes its effects on choice
between actions. It is clear, therefore, that it is the information that the stimulus
provides with respect to a predicted outcome that is critical to the ability of
such stimuli to affect performance, rather than its ability to evoke the condi-
tioned response. Nor, as it turns out, is this effect due to the ability of the stimu-
lus to retrieve and activate goal-directed control generally via retrieval of the
action–outcome relationship and, subsequently, the value of the outcome. One
of the more interesting and telling effects in this literature is the finding that
devaluing the outcome predicted by the stimulus does not affect the ability of
that stimulus to influence choice (Rescorla 1994; Holland 2004). Although the
predictive validity of the stimulus is critical, the value of the outcome predicted
by that stimulus is not.
Finally, it is important to note that these effects of predictive learning on
choice between actions are mediated by both excitatory and inhibitory action–
outcome associations and their effects on performance. When a stimulus pre-
dicting a particular outcome elevates the performance of an action associated
with that outcome, it does so without affecting the performance of other actions
(Laurent and Balleine 2015). We have hypothesized that this is at least partly
due to the fact that, in goal-directed learning situations, actions can become as-
sociated with the outcome that they deliver (e.g., A1 → O1, A2 → O2) as well
as with the absence of outcomes that they do not deliver (i.e., A1 → no O2,
A2 → no O1). Hence, a stimulus that predicts a particular outcome is likely
to elevate the performance of actions associated with the predicted outcome,
but not the responses associated with the absence of that outcome; informa-
tion (e.g., S1 predicts O1) can elevate an action, R1, that delivers O1 but not
another action, R2, that predicts no O1 (Laurent and Balleine 2015). In sum-
mary, Pavlovian-instrumental transfer provides insight into the way predictive
learning affects instrumental performance. As a phenomenon, it reveals the
following:
1. Presenting cues that predict specific outcomes elevates the performance
of actions associated with those outcomes (without affecting those that
are not) by selecting those actions and increasing (or inhibiting) their
execution in an ongoing manner.
2. The ability of stimuli to produce these effects depends on how specifi-
cally they provide information about those outcomes.
3. The ability of these cues to provoke the performance of actions does
not depend on the value of the outcome with which they are associated.
Of the theories advanced thus far to explain the effects of predictive learning
on instrumental performance, the one best supported suggests that the asso-
ciation between the stimulus and outcome, acquired during Pavlovian con-
ditioning, allows the stimulus to subsequently retrieve the outcome, thereby
retrieving (or inhibiting) the action associated with that outcome (Figure
3.2b; Balleine and Ostlund 2007). Any failure to inhibit competing actions
in the presence of a specific predictive stimulus would result in an unwanted,
intrusive action. This is strongly reminiscent of the occurrence of intrusive,
compulsive actions in various types of addiction as well as in multiple, se-
vere neuropsychiatric disorders, including Tourette syndrome (Leckman et al.
2010), grooming disorders (e.g., skin picking, trichotillomania; Chamberlain
et al. 2009), and obsessive-compulsive disorder (OCD) (Robbins et al. 2019).
Furthermore, not only does the superficial similarity of the behavioral influ-
ence of predictive learning on instrumental performance suggest this, there
is a very close similarity between the neural bases of Pavlovian-instrumental
transfer and the circuitry previously implicated in the compulsive actions as-
sociated with these conditions.
Neural Bases of Pavlovian-Instrumental Transfer
A number of structures have been found to be involved in the way predictive

learning affects action selection in transfer situations; for a summary, see Figure
3.2c. Many of these have also been implicated in either Pavlovian condition-
ing and/or instrumental conditioning, as would be anticipated. Treatments, for
example, that lesion or block the activity of the basolateral amygdala (BLA)
are particularly effective in blocking transfer effects, but they also abolish the
ability of animals to encode the association of Pavlovian cues and instrumen-
tal actions with specific outcomes. Similarly, lesions to or inactivation of the
mediodorsal thalamus (mdT) or the ventrolateral orbitofrontal cortex (vlOFC)
abolish the selectivity of Pavlovian predictions with respect to specific out-
comes, thereby abolishing the selectivity of transfer effects with respect to those
outcomes (reviewed in Cartoni et al. 2016). However, the projection from the
BLA to the nucleus accumbens shell is different in this regard: Disconnection
of the BLA from the shell abolishes Pavlovian-instrumental transfer but has no
detectable effect on either instrumental- or Pavlovian-conditioning processes.
This suggests that this circuit is critical for the way information derived from
predictive learning informs choice between instrumental actions (Laurent et
al. 2015a).
In recent years, considerable detail has been added to this basic circuit in
understanding how exactly the shell is involved in transfer. Briefly, what has
become clear is that inputs to the shell from the ventral tegmental dopamine
neurons are critically involved (Laurent et al. 2014) as are two other modula-
tory processes that are localized to the shell itself, involving the interaction
of opioid and cholinergic processes (Laurent et al. 2012, 2014, 2015b). The
34 B. W. Balleine
latter process is controlled by the giant aspiny cholinergic interneurons that

make up about 3% of the neurons in the shell. These neurons express delta
opioid receptors on their membrane which bind enkephalin that inhibits the
release of acetylcholine, thereby releasing its inhibitory effects on dopamine
D1 receptor-expressing SPNs in the shell (Bertran-Gonzalez et al. 2012).
Importantly, the degree to which delta opioid receptors are expressed on cho-
linergic interneurons is related to the effectiveness of predictive learning.
Mice that show strong Pavlovian learning also show strong delta receptor ex-
pression, and this is true whether one assesses excitatory (Bertran-Gonzalez
et al. 2012) or inhibitory associations (Laurent et al. 2015b).
The shell projects to the broader basal ganglia circuitry through several
pathways, but particularly to the medial and lateral segments of the ventral
pallidum. In a number of studies, we found that the more critical projection
for transfer effects is to the ventral segment of the medial ventral pallidum
(VPm) based on the effects of disconnecting the shell from the VPm versus
the VP-l prior to the Pavlovian-instrumental transfer test (Leung and Balleine
2013). Shell inputs to the VPm drive the output neurons from this structure,
which project to multiple efferent structures including the ventral tegmental
area (VTA) and the mdT (Root et al. 2015). Again, using disconnection pro-
cedures, we found that the critical output for the transfer effect is to the mdT,
whereas the projection to the VTA appears to be important for response vigor
irrespective of the direction of that response (Leung and Balleine 2015). Thus,
although damage to both VTA and mdT projections completely removes the
transfer effect, damage to the VTA projection reduces the overall level of per-
formance, whereas damage to the mdT projection blocks the bias in perfor-
mance typically induced during the transfer test.
The VPm to mdT projection is GABAergic and its activity inhibits thalamic
output neurons (Lavín and Grace 1994). These mdT output neurons inhibit the
vlOFC (Alcaraz et al. 2016); thus, turning off this mdT projection disinhibits the
vlOFC and alters the activity of this corticostriatal projection and its effects on
performance. As such, this loop serves the function of maintaining the selective
drive exerted by predictive learning on action selection, although precisely how
it influences motor performance is currently unknown. One possibility is that,
like other cortical areas that receive input from the thalamus, the OFC provides a
return projection that, ultimately, excites motor output through premotor cortex
(Hart et al. 2014). As this output is also likely to be engaged by the dorsome-
dial striatal goal-directed circuit, it is possible that the corticothalamic circuit
is generally functioning to allow predictive cues to potentiate the performance
of actions selected and evaluated by the goal-directed circuit (Hart et al. 2014).
Whatever the merits of these speculations, it is important to recognize that any
failure of the VPm to inhibit the mdT will result in a failure to disinhibit the OFC,
and that the effect of this failure is not to block transfer effects per se but to render
them nonselective; that is, the action typically inhibited in the transfer test via the
inhibitory response-outcome association will be disinhibited and released as an

intrusive, involuntary, or compulsive action.
Finally, one other structure implicated in the changes in instrumental per-
formance induced by predictive learning is the medial OFC (mOFC). Damage
to or inactivation of the mOFC also abolishes transfer effects without gener-
ally elevating performance; rather, Pavlovian cues have no influence on in-
strumental performance (Bradfield et al. 2015). One explanation for this effect
is that the mOFC critically mediates instrumental performance in extinction;
indeed we have recently argued that the mOFC functions to retrieve action
outcomes when they are freely recalled from memory as opposed to when they
are directly observable in the environment. We reached this conclusion from
the finding that mOFC manipulations altered sensitivity to outcome devalu-
ation, but only in tests conducted in extinction: when the devalued outcome
was delivered contingent on performance, the response was adjusted appro-
priately. Importantly, these manipulations also altered the animals’ sensitiv-
ity to Pavlovian-instrumental transfer (Bradfield et al. 2015). Thus, unlike the
vlOFC, the mOFC is not just involved in influencing predictive stimuli on
instrumental performance; it also affects experienced changes in reward on
that performance. This suggests that its effects on transfer might be related to
its involvement in retrieving the outcomes of goal-directed actions more gen-
erally when these outcomes are unobservable.
Neural Bases of Intrusive Thoughts and Actions
Over and above the relationship between the altered activity in various cortical-
dorsal-striatal circuits and the disease states described above, theories related
to the neural bases of addictive conditions and relapse as well as of OCD and
related conditions have implicated changes in the OFC-shell-VPm-mdT-OFC
loop mediating Pavlovian-instrumental transfer in those conditions (Fettes et
al. 2017). The role of this circuitry in OCD is particularly well documented,
especially the role of the vlOFC and mOFC and its reciprocal connections
with the mdT in this condition. Thus, for example, it has long been known that
the symptoms associated with OCD are ameliorated by surgical ablation of
either the OFC or of the midline and intralaminar thalamus, including the mdT.
Importantly, they are similarly ameliorated by thermal or gamma radio lesions
of the ventral anterior limb of the internal capsule (also called anterior capsu-
lotomy)—the region through which the white matter tracks pass containing
the bidirectional fibers which connect the thalamus and the OFC (Pepper et al.
2015). More recently, deep brain stimulation of these fiber pathways has been
reported to have a similar effect (reviewed in Greenberg et al. 2010).
Neuroimaging studies of OCD patients have implicated aspects of this cir-
cuitry in OCD, particularly the subregions of the OFC. Generally, the vlOFC
has been reported to be hypoactive (Remijnse et al. 2006) and the mOFC to be
hyperactive (Gillan et al. 2015) during tests that engage these regions; these
36 B. W. Balleine
findings are broadly consistent with other studies (e.g., Kahnt et al. 2012; Zald
et al. 2014). Two important findings from the animal literature suggest that these
differences in activity may be important for aspects of symptom formation. For
example, a study by Burguiere et al. (2013), which uses the SAPAP3 mouse
model of OCD, found that these mice show a persistent conditioned grooming
response when it was elicited by a cue that had been paired with a water drop to
the head (Burguiere et al. 2013). The authors hypothesized that this effect was
generated by a hypoactive vlOFC and thus sought to activate this region in an
attempt to reduce the excessive grooming response. They achieved this using
optogenetic stimulation of the vlOFC and found that the persistent grooming
was, indeed, inhibited by optogenetic stimulation of the lateral OFC neurons
projecting to the striatum. A similar study by Ahmari et al. (2013), optically
stimulated the terminals of mOFC cells expressing channelrhodopsin in the ven-
tromedial striatum, centering on the junction between the nucleus accumbens
core and shell. Stimulation of this pathway progressively increased grooming,
which persisted after the stimulation (Ahmari et al. 2013). As such it appears
that enhancing activation of accumbens projecting neurons in the vlOFC re-
duces compulsive grooming, whereas stimulating the mOFC projections to the
ventral striatum enhances compulsive grooming in line with the hypo- and hy-
peractive phenotypes of the vlOFC and mOFC in OCD, respectively.
Sometime ago, based on similar kinds of observations in OCD patients,
Modell et al. (1989) developed a circuitry model of OCD with striking simi-
larities to the circuitry involved in Pavlovian-instrumental transfer (Figure
3.2d). They argued that “the primary pathogenetic mechanism of OCD lies in
dysregulation of a basal ganglia/limbic striatal circuit that modulates neuronal
activity in and between posterior portions of the orbitofrontal cortex and the
associated medial thalamic nuclei” (Modell et al. 1989:32). Furthermore, they
proposed that specifically the compulsive symptoms of OCD are associated
with aberrant activity in a positive feedback loop in the reciprocally excitatory
frontothalamic neurons, due to the loss of inhibitory input from the ventro-
medial portions of the striatum. Further, they postulated that: “the net result
of ventral striatal activation is increased (inhibitory) pallidothalamic output:
essentially the loop transduces excitatory input from the vlOFC into inhibi-
tory output to the thalamus which serves to modulate activity of the frontotha-
lamic circuit by means of an interposed negative feedback loop” (Modell et al.
1989:32).
The same circuit is crucial to normal Pavlovian-instrumental transfer.
Damage to this circuit results in aberrant transfer; that is, it results in pre-
dictive learning eliciting both the performance of actions associated with the
outcomes predicted by those stimuli as well as actions that predict the absence
of those outcomes, which is the definition of an intrusive involuntary and mal-
adaptive movement. This hypothesis suggests, therefore, that the transfer ef-
fect and its attendant circuitry, which provides such an important translational
model of the cognitive control of actions, also provides a model of OCD in
which abnormality in this circuit results in the inhibition of the vlOFC and, by
extension, excitation of the mOFC.
Summary and Conclusions
Corticostriatal circuits play a central role in the induction of intrusive thoughts

and actions that are associated with major psychiatric disorders and addictions
of various kinds. Here, I point to two types of involvement based on an assess-
ment of (a) the corticostriatal circuits engaged during the acquisition and per-
formance of goal-directed and habitual actions and (b) the modulation of such
actions by predictive learning. The first indicates deviations in the balance be-
tween the dual corticostriatal circuits that mediate goal-directed and habitual
action control and the competitive and cooperative relationship between them
that is required to generate adaptive behavior. Changes to the goal-directed
circuit reduces its regulation of habits and allows the intrusion of stimulus-
driven actions without regard to their consequences. Alternatively, changes to
the habitual system could result in a loss of the cooperative processes through
which stimulus-driven urges provide the basis for the retrieval and execution
of goal-directed action. As a consequence, the operation of a planning process
that retrieves action–outcome associations could result in the mental simu-
lation of their performance without the ability to initiate those actions. This
simulation has much in common with intrusive thinking.
One initiator of this planning process is predictive learning. Stimuli associ-
ated with the outcomes with which actions are associated can directly elicit
those actions. Predictive learning can engage both excitatory and inhibitory
aspects of action control: excitatory in the sense that actions associated with
predicted outcomes are released, and inhibitory in the sense that such stimuli
can block this influence from extending to actions associated with unpredicted
outcomes. A circuit involving the orbitofrontal cortex, nucleus accumbens
shell, ventral pallidum, and mediodorsal thalamus controls this balance be-
tween the excitatory and inhibitory effects of predictive learning. Interference
with this circuit results in the release of inhibitory control, causing predictive
cues to elicit actions associated with unpredicted outcomes, which, to a first
approximation, is the definition of intrusive or compulsive action. Generally,
therefore, the arguments presented here suggest that it is not a failure of ha-
bitual or goal-directed control processes per se, but rather the failure of the
cooperative and competitive interactions between these processes that forms
the basis of intrusive thoughts and actions.
Acknowledgments
Support for this work was provided by a Senior Principal Research Fellowship from
the National Health and Medical Research Council of Australia (GNT#1079561). The
38 B. W. Balleine
author thanks members of the Decision Neuroscience Lab at UNSW and attendees at
the ES Forum on Intrusive Thinking for the many discussions that have helped to shape
this work.
4
What Is the State-of-the-Art

Toolbox for How Circuits
Mediate Behavior?
Michael R. Bruchas
Abstract
A critical challenge in the field of neuroscience as well as research into the neurobio-
logical basis of behavior has been to establish links between the cellular and biochemi-
cal processes within the brain and nervous system that occur during the mediation of
behavioral events. Developments over the last 10–15 years have provided several new
means to accelerate, advance, and dissect the specific mechanisms for brain function.
Developments across two realms of neuroscience and engineering have afforded re-
searchers, clinicians, and biologists advanced abilities to facilitate the dissection, ob-
servation, control, and perturbation of neural systems within intact, behaving animals.
These advances include electrical, optical, pharmacological, and specialized hardware
which allow for closed-loop interfaces to monitor and manipulate neural function. This
chapter explores how these recent developments have become integrated into our neu-
robiological tool chest. It describes current advanced approaches, and the limitations of
each, and explores future pathways toward even better technologies needed to dissect
the molecular, cellular, and circuit basis of behavior.
Introduction
The mammalian nervous system evolved over millions of years and contains a
heterogenous composition of networks and cells, which send messages to one
another to communicate information. This information is communicated by
a variety of signals: electrical, chemical, and anatomical (i.e., architectural).
These signals converge, amplify, or inhibit the flow of information to influ-
ence ultimately behavior in the organism. In many cases, specific central and
peripheral nervous system diseases are caused by dysfunctions in these brain
processes at molecular, cellular, and circuit-based levels. Some of these neu-
ropsychiatric disorders include, but are not limited to, addiction, pain, and
40 M. R. Bruchas
emotional disorders, such as anxiety, depression, or schizophrenia. For many

generations, neuroscientists have sought to understand the neurobiological ba-
sis of behavior with the specific intent of resolving and better treating these
types of disorders. Most current treatments rely on pharmacological or behav-
ioral modifications from trained clinical practitioners. In many cases, treat-
ments do not exist or individuals are resilient to any form of intervention.
A better understanding of the mechanistic underpinnings of behavioral pro-
cesses is thus needed to develop new methods to engage and hopefully adjust
the brain and spinal cord’s function toward a more typical homeostatic state.
The complexity of the brain and spinal cord’s inner workings has limited the
development of new therapies, but the technologies developed over the last
10–15 years offer promise in terms of uncovering these details, which in turn
could identify new targets or methods.
Furthermore, there is a growing interest in the neurobiological basis of
behavior as it relates to developing artificial intelligence, brain–machine in-
terfaces, and alternative methods for generating complex processor-based
systems and/or methods for adjusting human behavior in pathological states.
New probes are being developed which allow for the delivery and recording
of neuronal signatures in behaving animals. These technologies will provide
enhanced functionality for researchers, but they also open up new avenues in
clinical realms.
For neuroscience research to examine naturalistic and pathological behav-
ioral states, it must address a key challenge: the full integration of minimally
invasive, biological sensors, actuators, and pharmacological interventions.
This technology is advancing at a rapid, exciting pace, and many new ap-
proaches have started to become available. Here, I focus on these advances
(i.e., where we are in the field as well as the future pipeline of neurotechnolo-
gies) and discuss limits, ideas, and concepts for both biological and hardware-
based neurotechnology.
Observing, Recording, and Manipulation

of Neuronal Function in Vivo
The architecture of the central and peripheral nervous system is composed of a

series of integrated modules that span the molecular, cellular, circuit, and sys-
tem’s levels. At the basic molecular level, neurons and glia in the brain express
a selected series of proteins, which include ion channels (including iontropic
receptors), peptides, pumps, and G-protein-coupled receptors (GPCRs). In the
context of cellular homogeneity and heterogeneity, the brain is composed of
millions of neurons, each of which tends to be enriched with various receptors,
transmitters, or proteins. Over the last several decades, neuroscientists have
generally worked to classify these types of cells and have begun to understand
their diversity in response, release properties, and connectivity. The diversity
State-of-the-Art Toolbox for How Circuits Mediate Behavior 41
of cell types in the brain and spinal cord, along with the various scales at
which the nervous system operates, have made the design and implementation
of various neurotechnologies and interfaces challenging. Here I will describe
several technologies currently employed to observe and record neuronal ac-
tivity in awake or head-fixed animals: how the technology is currently be-
ing used, as well as future iterations of related methodologies. Devices that
interface with the nervous system must have the ability to decipher signals at
multiple scales as well as sufficiently interface with biological tissue in a flex-
ible, noninvasive manner.
Electrophysiological Methods
For nearly a century, the electrical properties of the brain and nervous sys-
tem have been empirically measured and recorded. Investigators have used
electrical probes to stimulate and mimic neuronal activity patterns in in-
vestigations of neural circuit functions during behavior. For many years,
researchers were severely limited by the “channel count” of the electro-
physiological probes for in vivo measures. This meant that we could not
sufficiently sample large populations of neurons in behaving animals. More
recently advanced materials engineering and manufacturing approaches
have, however, brought forth new technologies, including large Utah ar-
rays and Neuropixels (Jun et al. 2017). This latter technique integrates 960
recording locations within a 70 × 20 μm2 area and allows for single unit
action potentials to be isolated at very high resolution across multiple brain
regions. In addition, it reportedly allows for stable tracking of single neuron
activity over multiple days, thus allowing investigators to measure dynamic
changes within neural circuits that change over time. This method permits
significantly advanced throughput in neuronal recordings and provides the
ability, given their small size, for simultaneous recordings of activity to be
made over a wide range of brain regions.
Recent advances have also worked toward developing neural probes that
are softer and more flexible in their ability to interface with the brain. Most
neural probes are composed of hard surfaces, including metal, glass, and
silicon semiconductors; these materials can activate the brain’s immune re-
sponse and lead to severe lesion, inflammation, and cell death. Recently, an
injectable platform composed of mesh-based electronics, including 16 chan-
nels of platinum recording and stimulating electrodes, has become available
(Zhou et al. 2017). This device was designed in a flexible, open manner to
facilitate better integration with surrounding brain tissue and has the ability
to isolate local field potentials as well as single units. It can also be operated
chronically in the rodent brain for many months. This method provides ad-
ditional surface and material structure that can interact with the brain in a
noninvasive manner, while providing high-resolution actuation and observa-
tion data sets.
42 M. R. Bruchas
Current Limitations
Limitations of these new technological recording methods include the

following:
• Neuropixel devices collect a large volume of data. Thus, due to the
channel count and multiple-site recordings, data processing can be
slow, time-consuming, and difficult.
• Integrating multiple streams of data over time while assuring for qual-
ity and meaningful data remains a significant challenge.
• Large teams of computational neuroscientists are required to ensure
that the data is sufficiently and carefully managed, processed, and fit
into a larger picture.
• These devices also do not include the means for optical coupling and
isolating of selected optical approaches, which might help to identify
the neuronal type being recorded.
Future iterations of the technique will need to employ streamlined data pro-
cessing pipelines (now underway, as I understand it) as well as full integration
with optical methods for applications (discussed below). In the case of flexible
probes and the architecture therein, there are some limits to the technique in
terms of channel count and integration with related optical approaches. Future
iterations, however, may be able to incorporate some modifications to include
these features.
Neuronal Actuators and Related Devices
Electrical recording offers precise, high-resolution information related to spe-

cific activity of regions, networks, and single cells in the behaving animal.
However, one key limitation of the method is that investigators cannot simply
rely on a unit’s electrical signature waveform to classify the neuron as a certain
type. Although some investigators do use this method, it is not considered de-
finitive in most subdisciplines of the field. For over a decade, neuroscience has
employed the use of genetics to specify particular cells and to manipulate these
cells by incorporating light-sensitive proteins in the continually developing
and evolving field of optogenetics (Lerner et al. 2016). The “workhorse” in the
field has been channelrhodopsin-2 (ChR2), which acts to depolarize neurons
upon activation. This tool is expressed in a genetically defined manner to al-
low for selected excitation of a given neuronal population, alongside experi-
ments related to spatiotemporal sufficiency of a given neuronal subgroup in a
particular behavioral output. In the field of neural circuits and behavior, along
with in vivo neurophysiology, ChR2 has been used in numerous experiments
to dissect the specific properties of various behaviors, including reward, deci-
sion making, addiction, pain, anxiety, depression, social interaction, feeding,
and other homeostatic processes. In recent years, variations in ChR2 have been
generated to provide powerful new abilities that further extend the function
and utility of the optical approach. These include channels with faster kinetics,
step-function properties, altered activation spectra (redshifted), and cellular lo-
calization (Yizhar et al. 2011; Klapoetke et al. 2014).
As these optogenetic stimulation tools became available, additional ap-
proaches were being developed to allow investigators to “silence neuronal
activity.” Here, the principal strategy was to develop photosensitive cation
channels which could act to hyperpolarize a neuron and thus significantly min-
imize and prevent that neuron from generating an action potential and “firing.”
These inhibitory opsins include pumps for protons hydrogen (H+) (called ar-
chaerhodopsin or bacteriopsin), sodium (Na+), and chloride (Cl–) (called halo-
rhodopsin) (Yizhar et al. 2011). Like ChR2, these proteins have been modified
to enhance function, expression, sensitivity, and wavelength, thus affording
investigators more advanced methods to manipulate neural circuits. The key
advantage to these inhibitory opsins is their ability to be harnessed for the
determination of how a genetically defined neuronal population is necessary
for a given behavioral event. The investigators can time lock activation of the
optical silencing method within a given group of neurons and observe the be-
havioral consequences of that particular manipulation in real time.
While the optical tools described above offer spatiotemporal and optical
control of neuronal activity through excitation or inhibition of given subsets
of neurons, due to the constraints of their binary impact and the fact that
their activation does not necessarily mimic naturalistic neuronal activity,
additional optogenetic approaches have been developed. These include, for
instance, methods for specifically manipulating cellular signaling, neuromod-
ulation, and gene expression. In particular, a newer approach was developed
whereby GPCR signaling (the primary mediator of neuromodulatory function
in the nervous system) can be mimicked in a cell type- and neural circuit-
specific manner. These chimeric optogenetic tools have been engineered us-
ing seven transmembrane-spanning opsins that contain the intracellular loops
and C-terminal tail of GPCRs, which are typically expressed within mamma-
lian neurons or glial cells. One of the first families of opto-XR receptors was
the adrenergic receptor system, whereby optically active beta-2 and alpha-1
adrenergic receptors were generated (Airan et al. 2009; Siuda et al. 2015).
Subsequently, a series of new opto-XRs, developed over the years, can acti-
vate a whole host of G-protein signaling pathways and neurons, including Gi,
Gq, and Gs among others (Spangler and Bruchas 2017).
Finally, several other photoactivatable proteins have recently been em-
ployed in cellular studies and are beginning to be used in vivo in brain tissue.
These optogenetic tools target the inhibition or activation of second mes-
senger cascades (for a review, see Wiegert et al. 2017). These tools regulate
downstream signaling using allosteric or proximity-based effects. They in-
corporate the use of flavoprotein domains, such as the light-oxygen-voltage
domain and cryptochromes. These flavoproteins generally fit into one of
44 M. R. Bruchas
three categories and act to initiate enzymatic activity, dimerize, or change

conformation, all in response to light (Spangler and Bruchas 2017). The ad-
vantage of these newer optogenetic systems is that they can selectively and
discretely target, in a spatiotemporally precise manner, specific intracellular
signaling components within intact neurons or at the systems level; this al-
lows investigators to probe how a specific cellular signaling, trafficking, or
physiological event can be potentially directly linked to a behavioral out-
come in real time.
Additional approaches, now widely used in behavioral neuroscience re-
search, include “chemogenetic actuator” tools. These biological tools allow for
selective modulation of GPCR signaling in specific tissues or cell types. Like
their complimentary optogenetic partners, chemogenetic tools are very easily
adapted to behavioral contexts. The most widely used of the chemogenetic
tools are GPCRs, which have been designed to respond to specific ligands and
couple to specific excitatory (Gq, Gs) and inhibitory G-protein linked neuronal
pathways. In most cases, as with opsins, investigators use these DREADD
(i.e., designer receptors exclusively activated by designer drugs) proteins by
combining a particular genetic method (animal, viral, or both) to introduce
the DREADD into a particular cell type. Through molecular evolution of the
human muscarinic or kappa-opioid receptor, Roth and his group engineered
a family of mutated receptors that are only activated by clozapine-N-oxide
(CNO) or other ligands such as salvinorin B and compound 21, among oth-
ers (e.g., Roth 2016). Currently, most investigators use the DREADD proteins
hM3Gq and hM4Gi to excite (enhance probability) or silence neuronal or
gliotransmission, respectively. These two proteins are typically introduced via
a specific viral method (e.g., an adeno-associated virus, lentivirus, or herpes
simplex virus) into particular cells in the nervous system, after which investi-
gators inject a systemic ligand to activate each receptor protein for the desired
effect. The activity of these receptors and the drug CNO typically peaks at
90 minutes. This approach, while providing cellular and network-level access,
lacks spatiotemporal precision compared to optical methods. Thus, it is some-
what more challenging to incorporate into particular systems-level models of a
given circuit’s role in behavior.
Additional methods are now in the process of being engineered and tested
in a variety of systems. These include the development of magnetically sensi-
tive proteins as well as proteins that are sensitive to high-frequency vibrations,
such as ultrasonic actuation (Stanley et al. 2015; Wheeler et al. 2016). These
developments remain controversial (Meister 2016) yet could offer, if success-
ful, a completely new noninvasive means to perturb and probe neural circuit
function.
The use of various optogenetic and chemogenetic tools within neurosci-
ence is continuing to expand on an almost daily basis. The key advances in
protein engineering, crystal structure, cryogenic electron microscopy, and
biochemistry are allowing for continued progress in this type of toolbox. The
development of various sensitivities, color activation spectra, and functions

coupled with novel hardware approaches will continue to advance the field
and allow for more nuanced physiological neuronal activity patterns to be uti-
lized. Furthermore, with the advances in biological substrates in optogenetics,
additional hardware-based neuronal probes have been developed which offer
advanced optogenetic targeting and function to decrease neuronal tissue dam-
age or allow for tether-free animal movement in more naturalistic settings.
These devices include new Michigan- and Utah-based optoelectronic probes
(Deisseroth and Schnitzer 2013), along with printed flexible microLEDs and
multifunctional polymer-based fibers. In some cases, devices can be powered
using near-field or radio frequency communication parameters for completely
untethered control of LED function, and these devices can also be used in most
common behavioral setups in a closed-loop manner (Shin et al. 2017).
Key Limitations
Although optogenetic and chemogenetic approaches have provided unprec-

edented new knowledge about neural circuit function as it relates to behavior,
they are constrained by several limiting functions. Of primary importance is
the fact that sufficiency experiments that rely on using DREADD or ChR2 (and
related opsins) utilize broad neuronal activation via light or chemical entity to
activate a genetically defined neuronal population. In the case of chemogenetic
hM3Gq-DREADD activation, neurons that express the receptor will respond
in unison to the drug application; this increases their firing all at once, in a
similar manner, as the drug is exposed to and binds the DREADD receptor
onto a given neuron. In a similar manner, ChR2-based (or comparable) optical
excitation results in photostimulation across the entire field of cells, whereby
light can reach, with sufficient power, through the tissue to depolarize the pop-
ulation of neurons. The problem with this approach is that we know from in
vivo electrophysiological studies and optical imaging approaches that neurons
do not typically respond in a monolithic synchronous fashion. Indeed, neurons
fire in patterned, stochastic ways to encode various behavioral responses. This
is a severe limitation of current DREADD-, fiber- and LED-based optogenetic
strategies utilized in behavioral neuroscience. The assumption is that by stimu-
lating or inhibiting all the neurons in a given region at once, we can mimic the
activity of the neural circuit as it relates to a particular behavior. This “bina-
rization” of neuronal activation in circuit neuroscience should be questioned
and resolved, so that we can better understand the discrete, heterogenous na-
ture of circuit encoding and behavior in the nervous system. There is, however,
some promise with respect to this limitation, and a few papers have recently
utilized and highlighted this new method (Packer et al. 2015; Jennings et al.
2009). By using spatial light modulator-based two-photon microscopy, inves-
tigators have been able to image particular neuronal activity patterns and then
“play back,” in a finite manner, those patterns using optical stimulation across
46 M. R. Bruchas
a single planer view. This approach is limited to a fixed neuronal number due
to laser power and coupling restraints; however, the approach suggests that
advanced optical holographic methods can be utilized to overcome the limits
of monolithic, synchronous optical manipulations.
Detection and Visualization of Neural

Circuit Function and Transmission
The complexity of neural circuits in the mammalian nervous system has been
a daunting task to dissect. Canonical approaches have utilized in vivo elec-
trophysiological methods to record activity of neurons and ensembles within
discrete brain regions during behavioral tasks in awake, freely moving, or
head-fixed animals. This method, used for many decades, has provided neu-
roscientists with a rich framework to understand how various networks in the
brain respond under various behavioral states. While extracellular and multi-
dimensional (high channel count) recordings have been instrumental to our
understanding of neural circuits, they have been limited, for instance, by the
following factors:
• They lack genetic or cell-type identification.
• They are unable to track neurons across days and trials with confidence,
due to limitations of maintaining a single neuron during recording over
multiple sessions.
• Channel and cell count are limited by region and array size, and for
deep brain, significant issues arise due to lesioning of more dorsal
structures in an attempt to reach limbic structures.
Fortunately, several modern approaches have advanced our ability to detect
and visualize discrete neural circuits as well as to make claims about causality
of various cell types, circuits, and networks in mediating a particular behavior.
Although some of these approaches have only begun to be utilized widely in
the community, they are at the forefront of neural technology and are likely to
lead the field’s efforts in the coming years.
In vivo Calcium and Voltage Imaging
Advances in genetically encoded calcium indicators (GECIs) represent the

most recent avenue by which some of the limitations listed above are begin-
ning to be resolved. Older versions of GECIs were limited in their capacity
to resolve deep brain structures due to low signal, noise, and poor dynamic
range. Newer variants have helped to resolve many of these limitations (Chen
et al. 2013b; Odaka et al. 2014). The general principle of these GECIs is that
calcium ions enter the cell following an action potential, and these sensors
are then used to detect subtle changes in neuronal calcium by converting that
signal into a fluorescent signal that can be measured using advanced micros-
copy approaches. This allows for a reliable proxy measure of action potential
firing patterns, along with synaptic calcium dynamics in real time.
The recent development of ultrasensitive protein calcium sensors,
GCaMP6.0 series, has been transformative for the field at large, because they
allow for very reliable detection of calcium transient activity and circuit activ-
ity in deeper structures when coupled with the advent and use of both cranial
window-based imaging and the gradient refractive index lens (GRIN), which
provides optical advantages for researchers using single-cell imaging methods
within deep brain limbic circuits. Cranial window-based imaging in multipho-
ton applications can provide larger fields of view, so that hundreds and even
thousands of neurons can be imaged in a single animal over multiple behav-
ioral sessions. New “mesoscope” microscopes are becoming available for this
very purpose and will greatly expand the field. The utilization of these new
biological and hardware-based imaging tools with fiber photometry as well as
single- and two-photon imaging permits reliable terminal field and single-cell
detection of calcium transient activity over single and multiple trials span-
ning days to weeks. Various hardware has been developed and optimized for
maximizing the types of behavioral experiments in which GCaMPs can be
used. Fiber photometry has gained substantial popularity in recent years due
to its relative ease of use. This method employs a simple fiber optic and pho-
tometer-based detection system and provides a computationally simple data
pipeline (Gunaydin et al. 2014) for measuring “bulk” calcium dynamics in a
given neuronal population, and ease of use in freely moving behavioral studies.
Further specificity of single-cell activity has been gained through the advent
of GRINs lens-containing mini-endoscopes (Ghosh et al. 2011; Barretto and
Schnitzer 2012); this allows for a less than 2 g microscope to be mounted to an
animal’s head, and a complementary metal oxide semiconductor image sensor
for high-speed detection of calcium transients in single cells. This mini-scope
approach allows for deep brain imaging during freely moving behavior, to-
gether with single-cell tracking over the course of multiple behavioral sessions
(Mukamel et al. 2009; Xia et al. 2017). Finally, two-photon imaging in head-
fixed, awake-behaving rodents allows for high-resolution long-term imaging
with either cranial windows or GRIN lens implants for deep brain applications.
This head-fixed two-photon imaging can be coupled with complex behavioral
tasks, including virtual reality systems and spherical treadmills with simulated
environments for increased complexity in both endoscopic and cranial win-
dow-based imaging platforms (Zhang et al. 2018; Jennings et al. 2019).
Additional tools are under development and being tested. These include an
expansion of the color range for calcium indicator protein sensors, such as red
fluorescent calcium indicators (RCaMP) and jRGECO (Akerboom et al. 2013).
The advantage of these additional sensors is that they will improve the imaging
depth within intact brain tissue, since near-infrared light scatters less through
tissue. The other advantage of these red indicators is the future ability to allow
48 M. R. Bruchas
for simultaneous multicolor imaging of diverse genetically or circuit-specific

neuronal populations. These newly fashioned indicators could also allow for
dual imaging of presynaptic axon terminals or coupling with other optoge-
netic actuator-based approaches. These indicators are currently limited by their
signal-to-noise and dynamic range, but multiple groups are working to resolve
and optimize these issues using forward genetic screens in high throughput
testing scenarios.
One disadvantage of these calcium-based sensors is that they only sense
calcium transient activity, not actual action potentials. That is, they resolve
activity on a scale of seconds, not milliseconds. Compared to extracellular re-
cordings and phototagging, this has caused some in the field to remain skeptical
of the advantages of genetically encoded calcium imaging methods. However,
recent advances in genetically encoded voltage indicators may hold some
promise in resolving this issue. The recent developments of Archon, QuasAr2,
and CheRiff allow for reliable voltage detection (Adam et al. 2019), which has
been validated by electrophysiological studies to match kinetics directly with
neuronal action potentials.
Imaging and Detecting Neurotransmitter and Neuromodulator Activity
A very recent and exciting development in biosensors for other biomolecules

and neurotransmitters has come to the forefront. For many years it has been
difficult to measure the actions of both fast neurotransmitters and neuromod-
ulators in real time, using genetically encoded tools, during freely moving
behavior. Classically, these molecules have been measured using microdial-
ysis-based methods, which afford pristine detection with mass spectrometry
but are limited by spatial and temporal resolution. Dialysis is also generally
unable to detect larger molecules (e.g., neuropeptides), although some recent
progress has been made with opioid detection (Al-Hasani et al. 2018). These
include probes for glutamate (SF-iGluSnFR), acetylcholine (GACh), glycine
(GlyFS), GABA (Chamelean), dopamine (dLight and GRAB-DA) (Patriarchi
et al. 2018; Sun et al. 2018), as well as norepinephrine (GRAB-NE) (Feng et
al. 2019). These sensors rely on technology based on fluorescent-resonance
energy transfer or they take advantage of circularly permutated green fluo-
rescent protein (cpGFP, also used in GCaMPs) fusion proteins within the
third intracellular loops of specific GPCRs or related coupling domains. This
allows the sensor to detect “binding” of the transmitter or modulator, and
thus reveals the presence of a substance in a given response within a region
or circuit. These new sensors are promising because coupled with modern
viral and genetic tools, we can selectively express the sensors in various
brain regions, cell type, and circuits, and record activity in discrete behaviors
in real time.
While these protein-based sensor approaches offer a significant amount
of promise in terms of our ability to dissect the role and function of specific
neurotransmitters and modulators in real time during freely moving and head-
fixed behavior, there are still several limitations, and advances will be forth-
coming. Currently available sensors are mostly directed toward the detection
of small molecules, including fast transmitters, cholines, and monoamines.
While efforts are ongoing to use the same technology to detect neuropeptides,
tool development has been limited due to the high-affinity nature by which
neuropeptides bind to receptors and the inclusion of the cpGFP into the third
intracellular loop of the GPCR. This is the same region of the receptor that
dictates G-protein coupling and high affinity binding. There are some promis-
ing developing neuropeptide GPCR-based sensors on this front, none of which
have yet been published or validated in vivo; however, if neuropeptides can be
detected in a meaningful and genetically defined manner, it will open a host of
very exciting possibilities in how neuropeptides coordinate neuromodulatory
function within neural circuits that mediate a variety of behaviors (e.g., stress,
anxiety, fear, addiction, and reward seeking).
These new transmitter sensors could open new avenues and enable us to
address long-standing questions about neurons that co-release fast transmitters
(e.g., GABA and glutamate) while simultaneously releasing monoamines and
neuropeptides. We may be able to answer important questions about whether
neuropeptides encode specific information on their own or in conjunction with
specific fast transmitters under certain circumstances. Furthermore, these types
of tools coupled with imaging would allow us to expand our understanding
into the molecular and cellular basis of organization within neuromodulatory
and neurotransmitter circuits. Are peptides active at specific locations, released
in dendrodendritic form in some cases, or just released in mass in a volumet-
ric manner? Although these types of biosensors hold great promise, further
enhancement of their quantum yield (signal to noise) and sensors for other in-
tracellular signaling molecules (including cAMP, kinases, and other cascades)
will be needed for in vivo systems-level experiments in awake-behaving ani-
mal studies.
Conclusion
Optical tools provide unique methodologies in our quest to dissect neural cir-
cuits associated with behavior. Equipped with novel biological tools as well
as new, less restrictive hardware and/or systems with higher resolution for im-
aging activity within neural circuits, investigators have been able to resolve
specific spatiotemporal properties of discrete cell types, neurotransmitters, and
neuromodulatory pathways in real time during discrete behavioral events. The
challenges posed by these new methods include their invasiveness, their lack of
temporal resolution (particularly with GECIs), as well as their dynamic range.
Hardware limitations, in terms of the imaging window, pose limitations, for
instance, on data stream management. As richer, high-resolution data becomes
50 M. R. Bruchas
available through these new approaches, deciphering the data and utilizing new
computational approaches becomes more imperative. Computational models,
in turn, are needed to handle the new data, thus posing a future challenge.
The advent of these new technologies begs the question as to whether any
of the tools described above might eventually pave the way to a better under-
standing of intrusive thoughts in animal models, and whether these could be
applied in clinical translation. From the discussions at this Forum, it is clear
that there are a variety of possible uses for novel tools in establishing causality
and translation—provided that some of the limitations can be overcome. These
include using closed-loop sensing of neuronal activity (GCaMP or other) and
optogenetic (ChR2 or halorhodopsin equivalents) or pharmacological manipu-
lations in a wireless setting. Real-time sensing during established behaviors
defined to represent “intrusive thoughts” across species, alongside real-time
feedback with optogenetic and pharmacological control, would establish cau-
sality and mechanisms for intrusive thoughts, at least in one sense. For ex-
ample, deep brain stimulation has been widely used in clinical settings for a
variety of neurological and psychiatric diseases, yet it has not been used in a
closed-loop setting, whereby certain neuronal signatures, biomarkers, or other
measures would be detected followed by a closed-loop infusion of a drug or
optical/electrical stimulation.
The approaches outlined here, including optogenetic, chemogenetic, and
electrical perturbation, could be amenable to these ideas if we could (a) mea-
sure signatures of intrusive thoughts that span particular brain regions with
particular biomarkers and (b) overcome the limitations of expressing viruses
in the human brain. Recent developments in retinal research and clinical trials
with adeno-associated viruses, along with other viral delivery methods and
many new hardware developments, could assist translational approaches in
the future.
Collaboration between cross-disciplinary computational neuroscientists,
biologists, psychologists, behaviorists, and clinical psychiatrists is of para-
mount importance and needs continual encouragement. Notwithstanding these
challenges, the range of tools in the neuroscience toolbox continues to grow,
offering innovative ways to resolve specific pathways, networks, and behav-
iors with increased granularity. Future efforts require specific focus on cross-
species corroboration, computation, and analysis.
5
Convergent Experimental
Systems for Dissecting
the Neurobiology of
Intrusive Thought
A Road Map
Shannon L. Gourley, Antonello Bonci,

Michael R. Bruchas, Shelly B. Flagel, Suzanne N. Haber,
Peter W. Kalivas, Amy L. Milton, Paul E. M. Phillips,
Marina R. Picciotto, and Jeremy K. Seamans
Abstract
Nonhuman experimental systems (also known as model organisms) are critical for un-
derstanding the neurobiology of intrusive thought. These model systems allow for the
ability to manipulate specific neurocircuits, neurotransmitters, neuromodulators, and
physiological and intracellular signaling events associated with behavioral markers that
may be linked to intrusive thought. They permit unparalleled control over the external
and genetic environments in ways and to degrees that are not possible in humans. Intru-
sive thought is an emergent property of multiple systems: emotional, cognitive, motor,
and autonomic/somatic. In an animal model, one can ask specific questions about these
systems and how they may be linked to, permit, or suppress intrusions. For example,
how are specific connections, neuromodulators, or cell types involved in each of these
systems, and how do they help form or maintain behaviors consistent with intrusive
thought? Are positive versus negative valences unbalanced? Are common systems
Group photos (top left to bottom right) Shannon Gourley, Antonello Bonci,
Suzanne Haber, Peter Kalivas, Amy Milton, Michael Bruchas, Shelly Flagel, Paul
Phillips, Shannon Gourley, Jeremy Seamans, Antonello Bonci, Suzanne Haber, Marina
Picciotto, Paul Phillips, Peter Kalivas, Amy Milton, Jeremy Seamans, Marina Picciotto,
Shannon Gourley and Antonello Bonci, Shelly Flagel, Michael Bruchas
54 S. L. Gourley et al.
hijacked by intrusive thought, agnostic to the valence or content of the thought? Resolv-
ing these issues could be transformative for the treatment of several neuropsychiatric
illnesses that are commonly characterized by intrusive thought. This chapter presents a
road map for studying the neural mechanisms underlying intrusive thought using non-
human experimental systems.
Introduction
Understanding the neurobiology of intrusive thought requires unfettered and

unrestricted access to the brain. Thus, one turns to nonhuman experimen-
tal systems (also known as model organisms) because they allow admission
to the brain as well as unparalleled control over the external and genetic
environments, employing technical and experimental strategies that are not
possible in humans. This access permits dissecting, quantifying, and manipu-
lating specific neurocircuits, neurotransmitters, neuromodulators, and physi-
ological and intracellular signaling events associated with behaviors. Once
we develop strategies to infer intrusive thought in nonhuman experimental
systems, several goals can be pursued, such as the identification of neuro-
circuits which are, or are not, associated with intrusive thought. Are distinct
subcircuits, neuromodulators, or cell types involved in forming or maintain-
ing intrusive thoughts with positive versus negative valence? Are common
systems hijacked by intrusive thought, agnostic to the valence or content
of the thought? Resolving these issues could be transformative in develop-
ing treatments for several neuropsychiatric illnesses that contain intrusive
thinking as a pathogenic endophenotype. As discussed at greater lengths at
other points in this volume, illnesses include common disorders such as drug
addiction, posttraumatic stress disorder (PTSD), depression, and obsessive-
compulsive disorder (OCD).
To develop a road map for studying intrusive thought in nonhuman experi-
mental systems, our discussion begins by defining intrusive thought in the con-
text of biological frameworks for the research laboratory. Next, we focus on
conceptualizing intrusive thought as an emergent property of multiple systems.
This leads us to formulate a road map for investigating intrusive thought in the
future. Finally, we conclude by exploring the point from which we started and
analyzing where we still need to go.
Defining Intrusive Thought in Biological

Frameworks for the Research Laboratory
Our first goal is to set forth principles by which we can capture aspects of
intrusions within the domain of experimental systems. Intrusive thought
has been defined as unwanted, unintended, conscious mental events lack-
ing control (Clark 2005). The aspects of this definition that we are best able
Systems for Dissecting the Neurobiology of Intrusive Thought 55
to capture, operationally and quantifiably, in an experimental subject are

neurobehavioral events that occur ectopically (e.g., out of their appropriate
contexts), recurrence, resistance to change, and induction of arousal. Such
events are insensitive to modification by external stimuli that would typi-
cally redirect neural or behavioral activity, and these intrusions often inter-
rupt adaptive behaviors.
Our aim is to develop a framework for how we might understand the neu-
robiology of intrusive thought using nonhuman experimental systems. A pri-
mary challenge is the inability of our subjects to express their thoughts, so
to speak, which forces us to interpret their behavior as a surrogate measure
of thoughts. To address this issue, we describe in Table 5.1 key concepts that
should optimize any given approach. These concepts include construct, predic-
tive, and face validities. Construct validity refers to the degree to which a given
experimental strategy accurately measures what it is meant to be measuring.
Predictive validity refers to the extent a strategy can make accurate predictions
about the human condition. For instance, if a drug has anxiolytic properties
in humans, it should have anxiolytic properties in a valid task of anxiety-like
behavior in a rodent or nonhuman primate. Finally, face validity refers to the
degree to which a given strategy reflects what it is attempting to model. We
might ask, “Does this approach seem like it will measure intrusive thought?”
Of course, reliability and reproducibility are also key considerations. In addi-
tion, we highlight the notion of antecedents, which, in this chapter, refers to
factors that predispose an organism to, or directly triggers, an intrusive thought.
Table 5.1 Validities and considerations in designing research strategies.
Construct validity The interpretability, meaningfulness, or explanatory power

of a given model; the degree to which a test measures what it
claims to be measuring: How well does it capture the underlying
constructs?
Predictive validity The ability of a model to lead to accurate predictions about
the human phenomenon: How well does a procedure identify
pharmacological agents tested in model organisms that have
therapeutic value in humans?
Face validity The extent to which a test is subjectively viewed as reflecting
the concept it intends to measure: Does it seem like it is really
going to measure intrusive thought?
Reliability Stability and consistency with which a variable of interest can
be measured; phenomenon is readily reproduced under similar
circumstances
Antecedents The extent to which conditions in the model recapitulate factors
that precede or trigger the phenomenon of interest (here, intru-
sive thought)
Conceptualizing Intrusive Thought as an

Emergent Property of Multiple Systems
Embedded in the argument that nonhuman experimental systems have utility

in studying the etiology and neurobiology of intrusive thought are two funda-
mental notions:
1. Intrusive thought is, to some degree, conserved across rodent and pri-
mate species (and thus likely has some adaptive origins).
2. Corollaries and consequences of intrusive thought can be measured and
quantified in the absence of speech.
To the first point, one can imagine instances in which multiple types of
organisms would benefit from uninterrupted and intensive thought, such
as when the goal is to escape from a predator. However, an individual
must also be able to modify and shift focus when the situation changes
(e.g., when the threat has been resolved) and engage in other behaviors
that are more adaptive or otherwise suited to present and evolving con-
texts. A failure to inhibit intrusive thinking can distract from achieving
adaptive goals.
As shown in Figure 5.1, we conceptualize intrusive thought to be an
emergent property of multiple systems: emotional, cognitive, motor, and
autonomic/somatic (for discussion of the origin of intrusive thought in these
systems, see Roberts et al., this volume). We envision a world represen-
tation that contains these four coexisting elements, which homeostatically
analyze and validate environmental and intrinsic (thoughts) stimuli to elicit
adaptive behavior. Emotional and motivating content draw on circuitry in
the central zone of Figure 5.1. Intrusive events trigger a deviation from
the homeostatic condition; thoughts contain more excessive motivational
and attentional relevance to the individual than is appropriate for the envi-
ronment. In neuropsychiatric pathologies characterized in part by intrusive
thoughts, this deviation is associated with a loss of proper regulation of the
inner circuitry by the outer circuitry, as indicated in Figure 5.1 (for defini-
tions of typical vs. intrusive thoughts envisioned by the model in Figure 5.1,
see Table 5.2).
We envision that any given mental health disorder can coopt different do-
main hierarchies. Identifying these hierarchies could offer clues into the neu-
rocircuits that one might explore in investigating etiologies and developing
treatment strategies. For example, in disorders in which cognitive behavioral
therapy can be effective, such as OCD, the cognitive domain plays a significant
role in generating overall circuit feedback that restores homeostasis and con-
trol of the intrusions. We hypothesize that the distinct disorders or endophe-
notypes of disorders defined by DSM-5 have the order of domain dominance
shown in Table 5.3.
Cognitive
Cognitive Thalamo-
Thalamo- cortical
cortical
ACC Insula ACC Insula

Emotional Motor
Emotional Motor
Accumbens Accumbens
Limbic Amygdala Striatum Amygdala
Intrusion Limbic Striatum
Autonomic
Somatic
Autonomic
Somatic
Figure 5.1 Intrusive thought is an emergent property of multiple systems. We propose

that emotional, cognitive, motor, and somatic domains are recruited to interpret sensory
or internal stimuli and generate adaptive responses. Navigating a complex world and
adjusting our behaviors appropriately requires homeostatic involvement of these com-
ponents. Depending on the arousal and motivation that emerges from this homeostatic
interpretation, the central region is recruited (illustrated by the central red circle) to aug-
ment motivational value and attention. Each domain contains circuitry within the cen-
tral region, as indicated by the brain nuclei listed (ACC, anterior cingulate cortex), that
contributes salience to adaptive homeostatic interpretations and responses. Normally,
we evaluate and modify our behavior, such as when in an aroused state that arises from
a threat. In assessing the threat, we iterate between the external and internal circuits
(bidirectional arrow), adjusting our appraisal of the stimuli through feedback between
the circuits to generate the most appropriate responses. Though random thoughts oc-
cur, they are continuously appraised and only become problematic when the appraisal
and motivation/arousal generated by the inner circuitry does not match the information
received from the outer circuitry. Accordingly, the inner motivational circuitry becomes
resistant to, or dominates, the outer cognitive circuitry (one-way black arrow from in-
ner to outer circuitry). This leads to a loss of homeostatic response and manifests as
excessive, maladaptive thoughts and possibly inappropriate behaviors. For instance, an
intrusion producing a pronounced feeling of anxiety that cannot be regulated in a non-
threatening situation will produce autonomic, emotional hyperarousal and inaccurate
conscious assessments of situations that could manifest as posttraumatic stress disorder.
Formulating a Road Map for Investigating Intrusive Thought
Keeping in mind the definitions, considerations, and concepts laid out above,
we can begin to develop a meaningful list of behavioral and other factors that
could be tractably measured in experimental systems (Table 5.4). For instance,
we could capitalize on the ability of an external stimulus to distract an experi-
mental subject from engaging in goal-directed actions:
• Will a rat in an operant-conditioning testing chamber respond for food
reinforcers even in the presence of an opioid-related cue?
Table 5.2 Definitions of a typical versus pathological thought intrusion.
Typical • Regulated by homeostatic cross talk between cognitive, emotional,

motor, and somatic domains (see Figure 5.1).
• Motivationally relevant information initially involves many classic
limbic circuits and brain regions (e.g., ventral prefrontal and orbital
cortices, insula, nucleus accumbens, and amygdala), see Figure 5.1,
which act in concert with the outer circuitry to validate and regulate
the state of motivation and arousal.
• Level of arousal and motivation is appropriately managed by the
outer circuitry, through feedback with the environment, to create and
modulate the behavioral response.
Pathological • Contains all of the elements of a normal event, except that the limbic
circuitry (purple area, Figure 5.1) is not properly managed by the cog-
nitive circuitry (outer circuit, Figure 5.1), thus creating an imbalance.
• The resulting state of motivated hyperarousal leads to stress that is
perpetuated without access to adaptive feedback and/or regulation by
the outer circuitry and environment.
• This imbalance can develop initially from any domain, and vari-
ous combinations may be more typical in different neuropsychiatric
disorders.
• The different domain hierarchies which create the intrusion predomi-
nant in a given disorder offer a potential focus for experimental explora-
tion into the underlying neuropathology of homeostatic loss of control.
Table 5.3 Utility of model organisms: behavioral measures relevant to intrusive

thought. Here we broadly summarize behaviors and additional considerations relevant
to investigations of intrusive thought in nonhuman experimental systems.
What can we measure?

• Disruption of goal-directed behavior (e.g., by drugs of abuse or aversive stimuli)
• Persistent avoidance of aversive stimuli (e.g., despite extinction conditions)
• Persistence of a given behavior, despite adverse consequences or punishment
• Persistence of a behavior in the absence of a conditioned stimulus or instrumental
contingency (e.g., conditioned freezing that generalizes or fails to extinguish)
• Cognitive domains that are known to be affected in human conditions
Additional factors:
• Vulnerability factors (e.g., early-life stress)
• Individual differences
• Behavioral comorbidities that model known comorbidities in humans
• Recurrence and potential worsening with time (akin to sensitization or kindling)
• Alternatively, will the rat instead attend to the cue at the expense of
goal-directed food seeking?
• What differs, at a neurobiological level, between the rat who becomes
distracted and the rat who stays on task?
Similar to food-seeking behavior, drug-seeking behavior can have goal-di-
rected properties, but the important measure that one might wish to collect in
the context of intrusive thought is the degree to which it competes with a pre-
sumably more adaptive behavior, such as food seeking in a calorie-restricted
organism.
Some recently developed procedures (a) force organisms to arbitrate be-
tween food-seeking behaviors and the avoidance of aversive stimuli and then
(b) track the extinction of avoidance behavior in the absence of the stimulus
(Bravo-Rivera et al. 2015; Rodriguez-Romaguera et al. 2016). Others measure
the reaction of organisms to uncertainty (d’Angelo et al. 2014, 2017; Eagle et
al. 2014; Morein-Zamir et al. 2018). Both strategies could be used to investi-
gate mechanisms of intrusive thought, particularly when considered with fac-
tors such as individual differences or antecedents to intrusive thought, many of
which can be recapitulated in the laboratory (Table 5.5).
Another type of intrusion that could be recapitulated in nonhuman exper-
imental systems is the sense of incompletion of a task that looms until the
process is completed. One example in humans draws from obsessive hand-
washing in OCD: the notion that one’s hands must be washed is all consuming,
generating hyperarousal and stress until one washes their hands, resolving the
intrusion. Procedures in nonhuman experimental systems, such as persistent
Table 5.4 Examples of potential domain hierarchies involved in disorders containing
maladaptive intrusions. The interacting domains defined in Figure 5.1 may be associ-
ated with particular neuropsychiatric conditions to a greater or lesser degree.
Condition Proposed Hierarchy of Domains

OCD Cognitive = motor > Emotional = somatic
PTSD Emotional = somatic > Cognitive = motor
Craving in substance use disorders Emotional = motor = somatic > cognitive
Rumination in depression Emotional = somatic = cognitive > motor
Table 5.5 Antecedents to intrusive thought can be recapitulated in the laboratory. For
arousal, long-term events are historical events that give rise to vulnerabilities and resil-
iencies, whereas short-term events refer to triggering factors.
Long-Term Events Short-Term Events

Early-life experiences (early-life stressors) Autonomic responses and stressors
Genetic correlates Conflict
Environmental insults (drugs of abuse, trauma) Emotional representations of envi-
ronmental stimuli
response to a drug-paired cue or the “observing response task” (see d’Angelo

et al. 2017), have utility in capturing both motor- and circuit-level aspects of
incompleteness. In addition, most instrumental behavioral tasks include a dis-
crete signal, such as a light, which designates the completion of a response re-
quirement. Omission of that signal generates extended responding, a behavior
that is potentially motivated by a sense of incompletion.
A third approach to studying the concept of incompleteness, which is not
mutually exclusive, would be to measure neural signals that demarcate the
completion of a response. For example, during cocaine self-administration, a
phasic dopamine signal is observed in the nucleus accumbens (NAc) of rats
upon completion of a response requirement (Phillips et al. 2003; Willuhn et
al. 2012). In some animals, this feedback signal becomes diminished and, as a
consequence, animals keep repeating the action, resulting in higher drug con-
sumption (Willuhn et al. 2014).
Five Strategies for Investigating Intrusive Thought
To study intrusive thought in nonhuman experimental organisms, we propose

five different approaches, summarized in Table 5.6 and discussed below.
Back-Translation, Susceptibility, and Resilience
It should go without saying that nonhuman experimental systems will be of

greatest value if used in conjunction with appropriate types of behavioral anal-
yses. There is, of course, inherent difficulty in translating clinical behaviors
directly into an animal “model” of a mental health disorder, and one can debate
whether it is even possible to model a mental health disorder in its entirety in
animals (Bale et al. 2019). Back-translation offers a complementary approach.
Broadly, this term refers to the identification of components of mental health
disorders in humans that can be measured in nonhuman experimental systems.
The concept of back-translation informs our first two categories: experimental
approaches are driven by (a) factors implicated in human behavior, and/or by
(b) vulnerability and resiliency factors that have been documented in humans.
Ultimately, of course, the aim is to have “simultaneous translation” with dif-
ferent research approaches in humans and nonhuman experimental systems
converging on the same findings (Milton and Holmes 2018).
Notably, one specific form of back-translation refers to the deconstruction
of mental health disorders into specific psychological processes that can be
studied in nonhuman experimental systems. In doing so, one should avoid
measures that are subjective and self-reported; instead, the research scientist
should look to data that are readily quantifiable and free of confounding influ-
ences. One example is performance on a battery of psychological tasks, such
as the Cambridge Neuropsychological Test Automated Battery. By identifying
deficits exhibited by patients (e.g., in attentional set shifting or working mem-

ory), it should then be possible to identify populations of animals that mani-
fest the same deficits. Population variation could be induced through genetic
manipulations, early-life experiences, more proximal life experiences (e.g.,
exposure to drugs of abuse), or perturbation of neural circuits or neurochemi-
cal systems. In other words, one could expose the nonhuman experimental
organism to putative vulnerability factors that would be expected to exacerbate
intrusive thought. The readily quantifiable nature of the deficits in psychologi-
cal processing should allow for identification of potential neural and neuro-
chemical etiologies as well as hypothesis testing using the full range of tools
available to animal researchers.
Treatment Mechanisms
Another strategy, which we call “treatment mechanisms,” refers to our under-

standing of why, biologically, certain treatments are effective. Cingulotomies
improve intrusive thought in chronic pain management by mitigating the dis-
tressing nature of pain, but not the pain itself. The mechanisms by which this
phenomenon occurs remain elusive and could hypothetically be examined in
nonhuman experimental systems, opening a window for identification and
deep interrogation of cells that are excited, inhibited, or otherwise modulated
by cingulotomy. Such cell populations could then be manipulated pharmaco-
logically, genetically, or through other strategies, such as those detailed by
Bruchas (this volume). This approach would allow one to isolate neurobiologi-
cal correlates of successful interventions whose identification could ultimately
point in the direction of new and better treatment strategies.
Decrypting the Ensemble
The recurrent nature of intrusive thoughts suggests that some type(s) of similarly
recurrent oscillatory or reverberating neural processes could be identified in ef-
fective models. These processes could then be exploited to expand understanding
into the etiology of recurrent thought, a strategy that we refer to as “decrypting
the ensemble.” This strategy is inspired, in part, by evidence that in subcortical
areas, the plasticity of conditioned fear-related behaviors is accompanied by tran-
sient, measurable changes in the expression of calcium-permeable and calcium-
impermeable α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA)
receptors (Clem and Huganir 2010; Rao-Ruiz et al. 2011). Additionally, changes
in the ratio of subtypes of N-methyl-D-aspartate (NMDA) receptors can deter-
mine whether a fear memory ensemble is stable or susceptible to strengthening
following plasticity (Holehonnur et al. 2016).
The specific content of intrusions can be common across a large number of
patients and interacts with the environment (e.g., an increase in the number of
patients reporting obsessions regarding contamination with acquired immune
Table 5.6 Approaches to empirically investigate aspects of intrusive thought in nonhuman experimental systems. These approaches are not mu-
62
tually exclusive. Each approach is named and briefly described. Each strategy could be combined with stimuli meant to trigger or exacerbate an
intrusive thought and/or procedures that introduce conflict. The motivation to pursue a specific strategy is described (i.e., the strengths of a given
approach) and examples of each strategy provided. These lists are not exhaustive and some examples cross categories.
Approach Motivation Examples

1. Back-translation investigates genetic/mo- • Informed directly by the human condition • Genetics: study the role of the Fmr1 gene in
lecular factors, circuits, and behaviors that (i.e., high construct validity) repetitive behavior in fragile X syndrome
are highly (often, causally) implicated in, • Circuits: investigate interactions between the
or comorbid with, human conditions prefrontal cortex and nucleus accumbens in
rodent addiction models, an approach origi-
nally inspired by clinical research
• Behavioral: investigate cognitive functions
disrupted in neuropsychiatric illness (e.g.,
attentional function)
2. Susceptibility and resilience investigate • Potential for informing precision medicine • Genetics: study known genetic risk factors in
known vulnerability and resilience factors • Understanding resilience could shed light on Tourette syndrome and addiction
in humans mechanisms of coping • Behavioral: identify neurobiological effects
of early-life stress or enriched environments
3. Treatment mechanisms investigates the • Informed directly by “what works and what • Surgical: deep brain stimulation, used to treat
neurobiology of effective treatments does not” in treatment multiple conditions, can be studied in nonhu-
• Could shed light on disease mechanism or man experimental systems
reveal new targets for treatment • Behavioral: conditioned fear extinction
procedures could recapitulate aspects of
cognitive behavioral therapy
• Pharmacological: use of buprenorphine treat-
ment for craving in opioid misuse
Table 5.6 (continued)
Approach Motivation Examples

4. Decrypting the ensemble decodes and • Causality and biomarkers can be simultane- • Use electrophysiology, imaging, or molecular
recreates neural representations associated ously identified labeling to identify cellular changes associ-
with events or stimuli not currently present. ated with a given behavior
• Thereafter, attempt to simulate the same be-
havior by recapitulating neural or molecular
changes
5. Quantify natural behaviors measures • Unbiased, exploratory, and comprehensive • Ultrasonic vocalization, heart rate variabil-
species-appropriate, ethologically relevant • Could benefit from within-subjects measures ity, physiological events (e.g., sleep, pupil
physiological events that occur at atypical to capture individual- and population-level dilation), clustering of naturally occurring
time points or frequencies. differences behaviors (e.g., feeding, grooming) that
manifest spontaneously in conjunction with
an external stimulus (e.g., foot shock, aggres-
sive encounter)
63
deficiency syndrome in the 1980s and 1990s). It is likely, therefore, that there
are particular concepts and associated neural ensembles that are recruited for
such intrusions. This phenomenon may reflect intrinsic differences in excit-
ability in specific neuronal populations, as is the case for amygdala neurons
which show high levels of cAMP response element binding protein (CREB)
phosphorylation (i.e., activation), which are subsequently more likely to be
recruited to fear-related neuronal ensembles following an aversive experience
than other neurons (Josselyn et al. 2001; Frankland and Josselyn 2014; Josselyn
and Frankland 2018). In other words, this subset of neurons with higher levels
of phosphorylated CREB are primed to respond to input. They therefore could
be an ensemble that could predispose an individual to an emergent (intrusive)
event. Similar phenomena have been described in the context of immediate
early gene expression (Suto et al. 2016; Whitaker et al. 2017), and these kinds
of strategies could be deployed formally to study intrusive thought.
Quantification of Naturally Occurring Behaviors

One way to minimize anthropomorphic bias in interpreting animals’ behavior is
to utilize a final approach, which we term “quantification of naturally occurring
behaviors.” Here, the investigator records species-appropriate, ethologically
relevant physiological events (e.g., ultrasonic vocalization or naturally occur-
ring grooming or feeding behaviors) and identifies neural correlates when these
behaviors deviate in quality or quantity from what is typical. In this and all of
our categories, the introduction of triggers (e.g., drugs of abuse or stressors)
will likely have utility in creating a situation that would cause intrusive thought,
which then manifests in a behavior that can be studied.
Considerations
It is important to acknowledge multiple limitations inherent in nonhuman ex-
perimental systems as well as in the strategies that we propose. These include,
but are not limited to, the following:
1. There is a risk that phenomena unrelated to intrusive thoughts are
measured.
2. False negatives may arise, due to an inability to collect sufficient in-
formation (e.g., in cases of multiple interacting brain regions and/or
neuromodulators).
3. Manipulations of molecules/cells using many current tools may not be
physiologically relevant (see Bruchas, this volume).
4. In experimental organisms, the emotional content of any given manipu-
lation is difficult to measure conclusively.
Utilizing multiple strategies should help us overcome these challenges and
identify convergences.
Another challenge that we anticipate is resolving the complexity of many

modern cellular/physiological data sets. A comprehensive brain map whereby
affective circuits are defined by key features at a complete reductionist level is
needed. Such a map would include:
1. RNA profiling of millions of neurons within the mammalian brain in
naive versus “intrusive event-like states” within discrete brain struc-
tures (e.g., Campbell et al. 2017; Saunders et al. 2018; Gouwens et al.
2019; Mickelsen et al. 2019).
2. Mapping a functional architecture of these cell types alongside dis-
crete behavior epochs or states using imaging and/or physiological
processes.
3. Environmental, genetic, behavioral, and pharmacological dissociation
of critical manipulations which impinge upon and alter the transcrip-
tional state and functional responsivity of these maps.
Finally, a fruitful conceptual strategy may be to switch from assigning special
functions to genes or neurons to computational perspectives of how ensembles
work together and coordinate complex behavior.
Naturally, behavioral data sets can also be quite complex. A fertile strategy
may be to differentiate learned association structures, based on the complexity
of the information that is stored to support that association, using the “model-
free” and “model-based” terminology from reinforcement learning. A model-
free computation assigns a single dimensional value to a stimulus based upon
the reliability of its association with a motivationally relevant outcome. In con-
trast, a model-based computation establishes a model of the environment that
can be used to explore potential and inferred connections between stimuli and
states. These concepts are discussed at greater length by Phillips and Milton
(this volume).
Where Did We Start, and Where Are We Going?
A primary goal of this chapter is to offer a road map for future investigations
of intrusive thought. In this final section, we describe a selection of relevant
published and ongoing investigations and consider how these investigations
may be further developed to interrogate intrusive thought in nonhuman experi-
mental systems.
Global Modulation of Networks and Brain Regions versus Individual

Parts of the Circuit: A Place In Between
Intrusive thought emerges from an interaction between several functional do-
mains—emotional, cognitive, motor (see Figure 5.1)—and is often triggered
by internal or external sensory stimuli. Typically, in response to stimuli, the
interactions between these domains flow smoothly, with each taking a leader-
ship role in the appropriate situation but then returning to the status quo. For
example, an internal or external stimulus might invoke fear, overwhelming
the individual until either an escape (movement) is executed or the cognitive
control system takes over if the fear is unwarranted. Mediating between these
functional domains requires the complex integration of information across
them to resolve the situation. Intrusive thought can be considered to be a con-
dition in which this mediation fails. Identifying the circuitry that underlies the
integration of information processing across the different domains is a first
step in understanding how and where these areas communicate, necessary to
developing new therapeutic targets.
Karl Wernicke first recognized that connectivity of brain structures, rather
than their locations, was the central feature of higher-order cognitive functions
(Wernicke 1885/1994). Expanding on the idea, Geschwind suggested that this
involves a combination of functional localization and connectivity, leading to
the idea that the brain is comprised of complex, interrelated functional net-
works (Geschwind 1965a, b; Catani and Ffytche 2005). Functional imaging
studies and graph theory techniques moved the field forward, demonstrating
large-scale distributed networks and the existence of nodes and hubs (Sporns
2011). A node is an area that is connected locally or connected within a func-
tional system. A hub is a node of a network that has unusually high connectiv-
ity to other nodes, or degree centrality, and high connectivity to other hubs, or
eigenvalue centrality (van den Heuvel and Sporns 2013). Hubs are thought to
represent regions for integrating and distributing information from multiple
cortical regions. They likely play an important role in cross-functional compu-
tational tasks, such as integrating limbic, cognitive, and motor control calcula-
tions for decision making.
The rostral anterior cingulate cortex (rACC) is a good candidate for con-
taining hubs, because it sits at the connectional intersection of the emotion,
cognition, and executive control networks. Indeed, the entire rACC is consid-
ered a hub of the brain’s global network (Buckner et al. 2009). However, the
region is large, and inputs from different prefrontal cortical (PFC) functional
domains vary across it. These connections could represent information pro-
cessing sequentially across subregions (i.e., from valuation to cognition to ac-
tion). Alternatively, a hub could be embedded within the rACC that integrates
information across them. Consistent with the literature (Morecraft and Tanji
2009; Morecraft et al. 2012), mapping the distribution and relative strength
of frontal cortical inputs across rACC in nonhuman primates reveals that PFC
inputs to the rACC follow three general gradients:
1. Ventromedial PFC (vmPFC) and frontal pole inputs are strongest in
the ventral and rostral parts of the rACC, with decreasing strengths in
dorsal and caudal regions.
2. Frontal eye fields and premotor areas inputs are strongest in the dorsal
and caudal regions, decreasing in rostral and ventral rACC regions.
3. Ventrolateral PFC (vlPFC) and dorsolateral PFC (dlPFC) inputs peak
in a more central position.
One region embedded within these gradients, however, receives inputs from
unexpected additional areas. In addition to inputs from expected connections
from cognitive control areas, the dlPFC and vlPFC, this region is also con-
nected with regions that are part of the emotional system, the orbitofrontal
cortex (OFC) and vmPFC, as well as with the sensorimotor system, the frontal
eye fields (Tang et al. 2019). Thus, this connectional hub within the rACC is
in a position to integrate information across emotional, cognitive, and senso-
rimotor systems. It is perhaps unsurprising, then, that both PTSD and major
depressive disorder show treatment response in an area in close proximity to
the rACC hub (Mayberg et al. 1997; Pizzagalli 2011; Chakrabarty et al. 2016).
The striatum is also an important structure for integrating and distributing
information. Although the striatum is classically divided into limbic, cognitive,
and motor regions, embedded within this general topography, terminals from
different frontal cortical areas interface in the rostral striatum, positioning it
optimally to contain hubs (Haber et al. 2006; Averbeck et al. 2014). Indeed, in
a specific location within the rostral caudate nucleus, terminal zones from the
inferior parietal lobule, an area important for perception, converge not only
with those from the dlPFC and vlPFC, as expected (Cavada and Goldman-
Rakic 1991; Yeterian and Pandya 1993), but also with projections from the
OFC and rACC. Thus, similar to the rACC, this hub combines inputs from
several functional domains. The connections of the rACC and striatal hubs are
examples of highly integrative, cross-functional regions with distinct combi-
national inputs that provide the anatomical substrate in which computations
about motivation, internal states, cognition, perception, and motor control are
linked to mediate adaptive behaviors based on the interaction of these func-
tions. Disconnection of these hubs will likely result in an imbalance between
goal-directed control, emotion, and higher cognition, and thus play a key role
in maintaining intrusive thoughts.
Seeking the Source of Switching
One important aspect of countering or managing intrusive thought is the abil-

ity to modify thoughts and actions, a key ingredient in adaptive responding to
external stimuli (Figure 5.1). Multiple structures discussed above are naturally
involved in these processes and could be a focus for future investigations, par-
ticularly those that enjoy considerable homology between rodent and primate
species. One such structure, the ventrolateral orbitofrontal cortex (vlOFC), has
been intensively investigated using an instrumental contingency degradation
procedure. In brief, the procedure requires nonhuman (or human) experimental
systems to generate operant responses for rewards such as food or juice. Then,
the experimenter modifies the likelihood that a given behavior will be rein-
forced, and organisms must update learned action–outcome associations to
modify their responding optimally. In a series of investigations, inactivation of
the vlOFC blocked the ability of mice to update response strategies (Gourley
et al. 2013a; Zimmermann et al. 2017, 2018; Whyte et al. 2019), consistent
with evidence that certain vlOFC neurons represent outcome-related memories
(Namboodiri et al. 2019).
The vlOFC interacts with aspects of the dorsal striatum, a key constitu-
ent of goal-directed action, to coordinate action–outcome response flexibil-
ity (Gourley et al. 2013a; Gremel and Costa 2013). Meanwhile, the use of
viral-mediated gene silencing and behavioral pharmacological strategies has
revealed several essential molecular factors within the vlOFC that optimize
its function. These factors include, but are likely not limited to, brain-derived
neurotrophic factor (Gourley et al. 2013a; Zimmermann et al. 2017; Pitts et
al. 2020) and its high-affinity receptor trkB (Pitts et al. 2018, 2020), Abl2 ki-
nase (DePoy et al. 2017), GABAAα1 receptor subunits (Swanson et al. 2015),
fragile X mental retardation protein (Whyte et al. 2019), and developmental
expression of integrin receptors (DePoy et al. 2019). These investigations
provide overwhelming evidence that the vlOFC is necessary for behavioral
switching, and they potentially shed light on molecular factors that are dis-
rupted when intrusive thoughts interfere with behavioral flexibility essential to
day-to-day function.
One common factor linking all of these proteins is that they regulate the
stability or turnover of dendritic spines, the primary sites of excitatory plas-
ticity in the brain. Whyte et al. (2019) revealed that updating expectations
regarding whether an action was likely to be rewarded reduced thin-type
dendritic spines, considered immature, on excitatory vlOFC neurons in mice.
Meanwhile, the proportion of mushroom-shaped spines, considered mature
and likely containing synapses, increased, potentially solidifying newly mod-
ified action–outcome associations to optimize future decision making and be-
havioral flexibility.
One function ascribed to the OFC as a whole is the updating of expecta-
tions, particularly under ambiguous circumstances; by extension, unbalancing
these connections via spine loss or inappropriate excitatory plasticity could
render expectations ambiguous and thereby vulnerable to intrusion by compet-
ing impulses. Consistent with these notions, exposure to cocaine (Gourley et
al. 2012a; DePoy et al. 2017; Pitts et al. 2020) and stress hormones eliminates
dendritic spines in the vlOFC, and identical procedures cause failures in the
action–outcome updating of stress hormones (Gourley et al. 2012b, 2013b;
Barfield et al. 2017; Barfield and Gourley 2019). Further, drugs that improve
behavioral updating appear to recruit local cytoskeletal regulatory systems
(DePoy et al. 2017). As a final note, artificially stimulating excitatory neurons
in this region also causes failures in action–outcome updating (Hinton et al.
2019), potentially by activating circuits associated with OCD. This idea is dis-
cussed at length by Balleine (this volume). Understanding the conditions under
which specific vlOFC connections are stimulated or quiescent could shed light
on how thoughts and actions can fail to be updated and become intrusive, and
how one recovers from their intrusion by switching cognitive strategies or be-
haviors, maintaining adaptive flexibility between emotional, cognitive, motor,
or somatic domains (Figure 5.1).
Unfortunately, cellular heterogeneity within several brain regions pertinent
to this discussion remains undefined. For instance, even within the striatum,
where cells subtypes can readily be distinguished based on dopamine recep-
tor constituents, dopamine D1 and D2 receptor-mediated neuronal ensemble
responses to cues and rewards over temporal dimensions, and as a function
of experience, remain opaque (for review, see Castro and Bruchas 2019).
Determining the contribution of each cell to learning, memory, stressor, and
drug reactivity, for instance, has previously been studied using in vivo physi-
ological approaches, yet the process of defining circuit- and cell-type specifics
in time and space is still in its infancy. As such, comprehensively understand-
ing the neurobiological bases of neuronal ensembles is a critical step if we are
to dissect and understand the aberrant patterns (signatures) by which intrusive
events occur.
Arousal Systems
Intrusive thinking includes arousal, and deviations from typical arousal states
can predispose one to, or acutely trigger, intrusive thoughts, a notion empha-
sized in Table 5.5. As such, understanding the mechanisms of arousal pres-
ents a point of entry into understanding intrusive thought itself. Within the
framework that arousal can decrease the threshold for permeation of intrusive
thoughts, we might consider the ability of acetylcholine (ACh) release elicited
by salient stimuli to alter the strength of signaling in thalamo-cortico-thalamic
loops, both acutely and persistently, via synaptic potentiation (Aramakis et
al. 2000; Kawai et al. 2007). This ACh-mediated elevation in activity may
alter the threshold for transmission of sensory information from subcortical
to cortical structures. The directionality of this signaling can vary across de-
velopment, with different ACh receptors mediating increases or decreases in
the transmission of sensory information (Aramakis et al. 2000; Heath and
Picciotto 2009).
Although we recognize that hallucinations may not reach a formal defini-
tion of intrusive thoughts, it could be useful to evaluate the particular circuits
for which we have direct evidence of a causal relationship with this perceived,
maladaptive mental event. Pharmacological studies (Warburton et al. 1985;
Fisher 1991), as well as evaluations of patients with loss of cholinergic neu-
rons (Dauwan et al. 2018), reveal that blocking muscarinic ACh receptors or
decreasing ACh levels in patients with Lewy body dementia (Tsunoda et al.
2018; Dudley et al. 2019) results in hallucinations. One proposal is that loss of
either nicotinic or muscarinic activity in corticothalamic circuits may underlie
these hallucinations (Esmaeeli et al. 2019). Another suggestion is that cortical
ACh increases the signal-to-noise ratio of perceived events, and that “mus-
carinic receptor activation in the cortex is involved in confining the contents
of the discrete self-reported conscious ‘stream’ ” (Perry and Perry 1995:240).
When cholinergic input to the cortex is lost, irrelevant sensory information
normally confined to subcortical circuits enters conscious awareness (Perry
and Perry 1995), and hallucinations result from “a failure of the metacognitive
skills involved in discriminating between self-generated and external sources
of information” (Kumar et al. 2009:119).
An additional set of studies suggests that there is a pervasive increase in
ACh levels throughout the brain in patients who are actively depressed: for
unipolar depression, see Saricicek et al. (2012); for bipolar depression, see
Hannestad et al. (2013). Elevated ACh may be a risk factor for depression
since remitted patients have intermediate ACh levels between actively de-
pressed individuals and healthy controls (Saricicek et al. 2012), as measured
by competition with a cholinergic ligand and validated by within-subject
challenge with a cholinesterase blocker (Esterlis et al. 2013). Relatedly, ACh
is a critical mediator of arousal and rapid eye movement sleep (Ma et al.
2018). At baseline, ACh input to the basolateral amygdala is very high, and
tonic activity of the cholinergic system can thus control both the level of
arousal of stress-related systems and the likelihood that a stressful event will
activate the basolateral amygdala (Picciotto et al. 2012). Currently, there is
no information on whether this increase in ACh levels is associated with
intrusive thoughts (e.g., rumination in depression), but this topic could guide
future experiments.
Manipulations and measurements of both cholinergic signaling and circuits
modulated by its receptors may represent a cross-species neurobiological ap-
proach ripe for translational evaluation. One consideration is that intrusive
thoughts can be represented in experimental settings by regulation of arousal
states along a multidimensional continuum. A possible dimension along this
continuum includes asynchronous, decoupled activity of the filter or gain do-
main which prohibits “normal” function in a given circuit, steering an organ-
ism toward a hyperaroused state. One particular example of this phenomenon
is found in the locus coeruleus, which projects broadly throughout the brain,
and its activity (tonic vs. phasic) is dictated by salience, context, and stress re-
sponsivity. The ability of the locus coeruleus noradrenergic system to dissoci-
ate attention signals from stressful ones depends on which inhibitory filters are
engaged. Along the temporal dimension, intrusive thought may have the effect
of dysregulating the inhibitory gain signal (typically regulated by neuromodu-
lators such as neuropeptides, monoamines, and steroids), thereby producing an
unwanted hyperaroused state.
Lessons from the Study of Cocaine
Nonhuman experimental systems of many species will self-administer drugs of

abuse, including cocaine, thereby providing a strong measure of face validity
to drug self-administration in animal models of human addiction. As in hu-
mans, drug seeking can intensify with time and experience, take on compulsive
properties, and persist despite adverse consequences, allowing for the intense
investigation of neurobiological etiologies.
Over the last 15 years, several research groups have modeled compulsive
drug-seeking behavior in the face of aversive consequences in preclinical mod-
els, with the ultimate goal of mimicking, as closely as possible, the symptoms,
diagnostic criteria, and features of addiction manifested by human patients
(Deroche-Gamonet et al. 2004; Belin et al. 2008; Economidou et al. 2009;
Marchant et al. 2014; Belin-Rauscent et al. 2016). In well-validated rodent
models, voluntary drug-taking and drug-seeking behaviors coincide with mild
foot shock punishments, which are used to create negative consequences fol-
lowing drug use. Exposure to foot shocks has revealed a divide in rodent phe-
notypes into two separate groups: (a) those which are shock sensitive, whereby
the rat ceases to press a lever after receiving the foot shock and (b) those which
are shock resistant, whereby the rat keeps pressing the lever despite receiving
the foot shock. Much like human populations who develop compulsive drug
abuse or addiction, approximately 30% of rodents exhibit the shock-resistant
phenotype.
The utility of the punished model of compulsive drug use in identifying
and developing translational therapeutics was demonstrated by Chen et al.
(2013a). In this study, the authors discovered that shock-resistant rats self-
administering cocaine show a marked reduction of activity in the prelimbic
PFC, a subregion of the PFC that is important in mediating behavioral flex-
ibility and decision making. When Chen et al. reversed hypoactivity of this
brain region via optogenetic activation of the prelimbic PFC, rats significantly,
and almost instantaneously, reduced their cocaine-seeking behaviors. These
findings led to clinical trials using a well-known, noninvasive form of brain
stimulation, repetitive transcranial magnetic stimulation (rTMS), previously
used as a treatment for depression. These clinical trials revealed that rTMS
reduces cocaine craving and cocaine intake, thus paving the way for larger
double-blind clinical trials that are the first to offer a promising treatment
against cocaine use disorder (Terraneo et al. 2016; Pettorruso et al. 2018). The
results of these translational studies highlight the importance of continuing
efforts in developing increasingly sophisticated rodent models of substance
use disorders, which are fundamental in leading to the next generation of
treatments against substance use and other addictions, and more broadly, in-
trusions on adaptive functioning.
The term “incubation” refers to progressive, time-dependent elevations
in drug craving and sensitivity to drug-related cues. Incubation is thought to
contribute to the maintenance and persistence of addiction and relapse (Lu et

al. 2004; Venniro et al. 2016). By studying the incubation of cocaine-seeking
behaviors in model organisms, we may also gain insight into certain forms
of intrusive thought. Craving is considered a key factor in triggering re-
lapse and can be triggered by drug-associated contexts or discrete drug cues.
Although craving in humans is typically triggered through a combination of
drug-associated contexts and cues, the two stimuli involve distinct and over-
lapping circuits. Thus, in animal models, it is useful to isolate them from each
other during the incubation period; for instance, by extinguishing behavioral
response to the drug-associated context with daily exposure to the context in
the absence of drug availability. Accordingly, when the drug-paired cue is re-
turned, it becomes possible to quantify behavior motivated only by the cue
and to evaluate changes in circuitry and cell physiology produced by the cue-
induced motivational state.
Within the context of dopamine-related pathologies (which presumably
include substance use disorders), a change in the configuration of dopamine
D3 receptors in the NAc has been observed, particularly when tonic dopa-
mine levels are low. This change has not yet been fully characterized but may
include a change in the ratio of the D3nf isoform; it also seems to enhance
functional coupling with dopamine D1 receptors. This change has been as-
sociated with ticks in Tourette syndrome (Frau et al. 2016), L-DOPA induced
dyskinesia (Fanni et al. 2019), and pathological gambling following dopamine
agonist treatment in Parkinson disease (Pes et al. 2017). Treatment with the
5α-reductase inhibitor, finasteride, reverses associated molecular changes and
ameliorates each of those pathological traits. Preliminary data indicate that
finasteride also reverses incubation of cocaine craving and reduces escalated
cocaine consumption (P. E. M. Phillips, pers. comm.), a finding that may be
relevant to intrusive thought as well.
Distinct drugs of abuse (e.g., cocaine, heroin, alcohol) induce both similar
and divergent neurobiological changes in brain regions like the NAc and fron-
tal cortex. In light of the overlap in drug-seeking endophenotypes produced
by, for example, the self-administration of opioids and psychostimulants, and
the shared vulnerability to relapse in addiction, one argument has been that
understanding the shared (rather than distinct) neurobiological factors might
be particularly fruitful. One such example is elevated synaptic glutamate
spillover from prelimbic PFC projections in the NAc during drug seeking
for cocaine, heroin, alcohol, and nicotine. Under typical conditions, synap-
tic glutamate spillover is moderated by the glial glutamate transporter, GLT-
1, located on astroglial end feet adjacent to the synaptic cleft. GLT-1 tightly
controls basal extracellular glutamate, protecting against synaptic glutamate
spillover. However, several drug classes downregulate GLT-1 and retract glial
end feet from NAc synapses, modifications that have been directly linked to
drug-seeking behavior (for a review, see Bobadilla et al. 2017).
Notably, acute stress has many overlapping effects, including decreased

glial synapse coverage and reduced GLT-1 in the NAc (Garcia-Keller et al.
2016) and induction of glutamate spillover from unprotected glutamatergic
synapses, which produces transient synaptic potentiation in NAc dopamine
D1 receptor-expressing neurons (Scofield et al. 2016). Transient potentiation
creates a situation that reinforces behavioral responses to stress- or drug-paired
cues, making responding more persistent and more competitive with other
stimuli (Kalivas and Kalivas 2016). Given that intrusive thought is similarly
characterized by recurrence and resistance to outside stimuli that would typi-
cally redirect behavior, we should directly test whether these factors contribute
to intrusive thought.
Sign Tracking, Goal Tracking, and the Interruption of Top-Down

Control over Behavior
When rats are exposed to a Pavlovian-conditioned approach procedure,

wherein an illuminated lever, a conditioned stimulus (CS), precedes the deliv-
ery of a food reward, an unconditioned stimulus (US), into an adjacent food
cup, two distinct phenotypes may emerge (Flagel et al. 2009): goal trackers
and sign trackers. Upon lever-CS presentation, goal trackers approach the lo-
cation of reward delivery, whereas sign trackers approach and interact with the
lever-CS itself.
For both sign trackers and goal trackers, the lever-CS is a predictor because
it elicits a conditioned response. For sign trackers, however, the lever-CS also
acquires incentive motivational value (incentive salience) and is transformed
into a “motivational magnet” (Berridge and Robinson 2003). That is, for sign
trackers, the lever-CS itself is attractive, elicits approach behavior, and acts as
a conditioned reinforcer (Cardinal et al. 2002; Berridge and Robinson 2003;
Flagel et al. 2009). Sign-tracking behavior can be considered compulsive be-
cause it will persist even if it results in omission of reward delivery, and is
resistant to extinction (Tomie 1996; Flagel et al. 2009; Ahrens et al. 2016).
Furthermore, relative to goal trackers, sign trackers are more impulsive (Lovic
et al. 2011): they exhibit deficits in sustained attention (Paolone et al. 2013),
show exaggerated responses to aversive stimuli (Morrow et al. 2011), and have
an increased propensity for cue-induced reinstatement of drug-seeking behav-
ior (Saunders and Robinson 2010). Recent evidence in rats (Eagle et al. 2014;
Vousden et al., submitted) and humans (Albertella et al. 2019) indicates that
sign trackers show greater levels of dysfunctional checking behavior of rel-
evance to OCD.
The sign-tracker/goal-tracker animal model has been used to parse the neu-
ral mechanisms underlying two different learning strategies: predictive versus
incentive learning. Sign tracking, or incentive learning, is dependent on do-
pamine in the NAc (Flagel et al. 2011b). In fact, using this model, it has been
shown that the shift in dopamine in the NAc from the reward (US) to the cue
(CS) encodes the incentive value of the cue, not the predictive value (Flagel et
al. 2011b). Relative to goal trackers, sign trackers show greater engagement of
the cortico-thalamic-striatal “motive circuit” in response to a food cue (Flagel
et al. 2011a). Within this circuit, the paraventricular nucleus of the thalamus
(PVT) has emerged as a critical regulator of individual differences in cue-moti-
vated behaviors (Haight and Flagel 2014; Haight et al. 2015; Kuhn et al. 2018).
The PVT is a midline thalamic nucleus ideally located to integrate cog-
nitive, emotional, and arousal information from various areas of the brain
and, in turn, to guide motivated behaviors (Kelley et al. 2005; Kirouac 2015).
Specifically, the PVT receives dense input from the PFC, as well as subcortical
areas, including brainstem nuclei such as the dorsal raphe and locus coeruleus
and other areas such as the lateral hypothalamus and amygdala. The PVT then
integrates this information and sends reciprocal output to some of the same
regions, but also has dense glutamatergic projections to the shell of the NAc.
In fact, the PVT can regulate dopamine release in the NAc, even in the absence
of the ventral tegmental area (Parsons et al. 2007).
Within the context of the sign-tracker/goal-tracker animal model, neu-
rons projecting from the prelimbic PFC to the PVT appear to encode the
predictive value of reward cues, whereas subcortical systems surrounding
the PVT encode the incentive value. Specifically, sign-tracking behavior is
thought to result from hyperactivity of neurons projecting from the lateral
hypothalamus to the PVT, and those projecting from the PVT to the NAc
(Haight et al. 2017). The working hypothesis, therefore, is that cognitive rep-
resentation, or the predictive value of the reward cue, is encoded in prelimbic
PVC-PVT projecting neurons, and that this top-down process predominates
in goal trackers. In sign trackers, however, where incentive learning prevails,
subcortical processes are able to override this top-down mechanism. Thus,
the PVT appears to act as a fulcrum between top-down cortical processes and
bottom-up subcortical processes, and an imbalance between these processes
may result in aberrant or psychopathological behavior. It is also intriguing, in
light of our discussion of ACh systems above, that sign-tracking behavior in
rats is associated with poor attentional control, mediated by an unresponsive
basal forebrain cholinergic system (Kucinski et al. 2018). The neurobehav-
ioral endophenotype of sign trackers may capture antecedents that predis-
pose an individual to intrusive thoughts. For example, sign trackers appear
to have an inherent imbalance between emotional and cognitive domains
(with the PVT acting as a fulcrum between the two domains; see Figure 5.1),
and this imbalance renders them more susceptible to behavioral control by
intrusive experiences.
Avenues for Future Research
The concepts outlined above provide a number of avenues for future research.
For example, the idea that we can capture and decode the neuronal ensembles
associated with disruptions of behavior and intrusive events suggests that we

can use those neuronal ensembles to test whether activation of neurons, path-
ways, brain areas, or interregional networks recapitulate or disrupt intrusive
events. A prototypical experiment might be to provoke an initial adaptive be-
havior, such as grooming in a rodent as a result of a sticky substance on its fur
(e.g., peanut butter). During the adaptive behavior, molecular techniques could
be used to capture active neurons that drive both the sensory representation of
the sticky substance and the motor output. Driving the activity of these neuronal
ensembles repeatedly should recapitulate the behavior and the sensory repre-
sentation, without the feedback of the cleaning of the fur or a clear outcome of
the motor behavior. Measuring the increasing connection between the sensory
and motor systems could be achieved with electrophysiology, or behaviorally,
by measuring the likelihood of grooming to other, less intrusive stimuli (i.e.,
stimuli that would not generally elicit a robust response). Generalization to
other stimuli could also be measured. Finally, recruitment of circuitry related
to anxiety or emotional behavior could be measured, which might result from
the mismatch between the environment and the sensory perception or motor
outcome. This hypothetical “peanut butter test” could be generalized to other
behaviors in which a specific initial stimulus and its adaptive motor outcome
are initially paired and then dysregulated, by driving the neuronal correlates of
the event and subsequent outcome in the absence of appropriate feedback (e.g.,
cued fear memories, drug-associated stimuli). Capturing aberrant outcomes of
neuronal recapitulation may identify common neural mechanisms, whether of
plasticity or systems-level generalization, which could then be probed in pa-
tients with intrusive thought.
Conclusions
Intrusive thoughts are a hallmark of several psychiatric conditions (e.g., OCD,

panic disorder, major depressive disorder, and addiction). For individuals suf-
fering from these disorders, intrusive thoughts gain inordinate control over
their emotions and actions, interfering with daily activities and disrupting
lives. Given that intrusive thoughts are common to multiple mental illnesses, it
is surprising how little we know about the underlying brain mechanisms. This
gap in knowledge stems from the fact that we have not yet, as a field, tried to
capture explicitly the commonalities of multiple disorders, such as intrusive
thoughts. Here, we have highlighted standard animal models, behavioral tests,
and outcome measures that could be exploited to shed light on the neurobio-
logical components of intrusive thought. We defined intrusive thought within
a biological framework that can be probed in the research laboratory, with
resultant models optimized to yield novel therapeutic targets. We proposed a
conceptual model that captures intrusive thoughts as an emergent property of
multiple systems (emotional, cognitive, motor, and autonomic/somatic) that
are represented in hubs throughout the brain. When the neural choreography
between these hubs and their corresponding nodes becomes disrupted, there
is a loss of homeostatic and/or cognitive control which leads to maladaptive
thoughts and inappropriate behaviors. In the laboratory, with careful experi-
mental design and multidimensional levels of analyses, we can model this loss
of control on both behavioral and neural levels. Here, we provided a road map
that illustrates multiple routes by which different approaches can be used in
combination to expand our understanding of intrusive thought. This road map
is not proscriptive; rather, we hope that it will serve as a foundation for a novel
avenue of preclinical research to advance our knowledge and ultimately lead
to more effective therapies for a number of psychiatric illnesses that are char-
acterized by intrusive thoughts.
Acknowledgments
The authors wish to acknowledge support from the United States National Institutes of
Health (grant numbers DA014242, DA044297, MH117103, MH007681, MH100023,
OD011132, DA039687, AA024599, MH106428-5877, DA033396, MH111520,
MH112355, MH045573, and MH106435).
6
How Can Intrusive Thinking

Be Measured in Vivo and
Studied in the Context of
Brain Mechanisms?
Marie T. Banich
Abstract
This chapter reviews different methods that can be used to examine and understand
intrusive thought, beginning with behavioral methods. Common among these are
self-report and diary measures of the experience, duration, and intensity of intrusive
thoughts as well as self-reports of the difficulty in controlling such thoughts. These
questionnaires, for the most part, have been tailored to the types of intrusions specific to
a given psychiatric syndrome (e.g., flashbacks in posttraumatic stress disorder, thoughts
of contamination in obsessive-compulsive disorder), which highlights the need to cre-
ate a transdiagnostic self-report measure. Another common behavioral paradigm is to
investigate intrusions after individuals are exposed to traumatic material, through a
symptom provocation paradigm in individuals who have experienced trauma or an ana-
log trauma (e.g., viewing a disturbing movie). Other behavioral paradigms, such as the
Think/No-Think paradigm, specifically examine mechanisms of memory retrieval and
suppression often thought to be disrupted in posttraumatic stress disorder.
Thereafter, it addresses paradigms for examining the neural mechanisms associ-
ated with intrusive thoughts. These approaches primarily couple behavioral techniques
or paradigms with functional magnetic resonance imaging or electroencephalographic
(EEG/ERP) methods. In addition to providing insights into the neural mechanisms that
may underlie intrusive thoughts, these approaches may provide additional information
regarding cognitive mechanisms, such as discerning whether memories are being sup-
pressed or replaced. Discussion concludes by examining emerging approaches to the
study of intrusive thinking. A main challenge is to find a method to verify that intrusive
thoughts have indeed occurred. New paradigms that combine neuroimaging techniques
with computational methods drawn from machine learning offer promise, as do tech-
niques which allow intrusive thought processes to be examined as they occur during
more naturalistic processing (e.g., watching a film).
80 M. T. Banich
Introduction
The question of how one can measure intrusive thoughts (i.e., thoughts that
“pop up” into consciousness in a seemingly uncontrolled manner) is difficult
to address, especially in a laboratory setting. It can be challenging to capture
and to verify their occurrence by measures other than self-report. Moreover,
traditional experimental measures used to explore mental processes (e.g., those
that determine reaction time and/or errors) are not applicable. Despite these
challenges, a number of approaches, both behavioral and biological, have
been employed. In this chapter, I review those methods, discuss some new
approaches, and suggest ways that the knowledge from other related arenas
of inquiry might be used to inform potential novel approaches to this difficult
question.
The need for scientists to have methodological approaches that will enable
an understanding of the basic cognitive and neural processes underlying in-
trusive thinking is apparent when one considers that intrusive thoughts are
ubiquitous across a large number of psychiatric disorders. Although intrusive
thoughts are generally associated with posttraumatic stress disorder (PTSD)
and are captured by the intrusions cluster of diagnostic criteria (American
Psychiatric Association 2013), they occur in many psychiatric disorders: de-
pression is characterized by intrusive negative thoughts and memory (Newby
and Moulds 2011), anxious apprehension (i.e., worry) by repetitive intrusive
concerns about future negative events (e.g., Fox et al. 2015), obsessive-com-
pulsive disorder (OCD) by thoughts of contamination and/or harm to self or
others (Bouvard et al. 2017), and schizophrenia by intrusions of semantic and
sensory information (Elua et al. 2012). This commonality raises the possibility
that intrusive thought across disorders may have a common underlying neural
circuitry (Kalivas and Kalivas 2016). What tends to vary somewhat across dis-
orders is the content of those intrusive thoughts (e.g., negative attributions of
the self in depression vs. concerns about potential dangerous future outcomes
in anxiety). In addition, intrusive thoughts in disorders such as PTSD are gen-
erally thought to be more sensorially based and of shorter duration, compared
to ruminative thoughts associated with depression and worry, which tend to
be more cognitive in nature, of longer duration, and recurrent (Speckens et
al. 2007). Nonetheless, what all these types of intrusive thoughts share is that
they appear to impinge upon consciousness in a somewhat uncontrolled man-
ner. Hence, my focus here is on methods for examining the complete range
of types of intrusive thoughts. To date, however, the majority of the work us-
ing methods to examine intrusive thought has focused mainly on individuals
with PTSD, OCD, and/or nonclinical populations. Thus, there is a much larger
range of individuals with psychiatric disorders to whom such techniques might
be applied.
Here, I will consider a number of different ways in which intrusive thoughts
have been examined. I begin with a discussion of behavioral paradigms, from
Intrusive Thinking in Vivo and in the Context of Brain Mechanisms 81
questionnaires to experimental procedures, that examine how the frequency,

nature, and control over intrusive thoughts can be measured. Thereafter, I con-
sider how different brain-based techniques can be used to shed light on the
underlying nature and neural bases of intrusive thought and conclude with a
discussion of how intrusive thoughts might be verified in ways other than by
self-report, specifically looking at emerging techniques that examine or track
their representational content.
Behavioral Methods: Measuring the Frequency,

Nature, and Control over Intrusive Thoughts
Self-Report and Questionnaires
Foundational work to understand intrusive thought comes from behavioral ap-

proaches that rely on self-report, of which there are two main types: The first
tries to assess the content and nature of intrusive thoughts, while the second
assesses the ability to control or manage thoughts, both those that are intrusive
and other thoughts more generally. In addition, there is a third type of self-
report that is a hybrid, assessing both the nature and controllability of thoughts.
Questionnaires that examine thought content have generally been designed for
specific psychiatric diagnoses. A number of these measures and their charac-
teristics are outlined in Table 6.1.
For the most part, these measures appear to have good reliability and va-
lidity, so that scores on these scales can be used (or combined) in studies that
take an individual differences approach in examining how scores on these
measures vary with other metrics of interest. For example, scores on these
measures can serve as covariates in brain-based approaches to determine
whether the degree of brain activation in particular regions is associated with
increasing scores on such measures. This approach is discussed in more de-
tail below (see section, Examining the Neural Systems Related to Intrusive
Thoughts).
Importantly, research using these measures indicates that most individuals
experience intrusions and that intrusive thoughts occur across a continuum
from nonclinical to clinical populations; for example, a similar factor structure
is observed in both groups (e.g., Reynolds and Wells 1999). Factors that may
distinguish clinical and nonclinical populations include the frequency, impor-
tance, and difficulty in managing those intrusions (e.g., Clark et al. 2014a).
These findings are notable as they suggest that new and emerging methods used
to examine intrusive thoughts in nonclinical populations could be fruitfully
employed within psychiatric populations. For example, research with nonclini-
cal populations has attempted to isolate the basic dimensions that underlie re-
petitive (although not necessarily intrusive) thoughts by asking individuals to
rate the nature of their most common thoughts on a number of dimensions,
82 M. T. Banich
Table 6.1 Examples of typical self-report measures of intrusive thoughts and the abil-
ity to control thought. Typical self-report measures of intrusive thoughts and the ability
to control thought. Example items from these measures are shown in italics.
Name Construct Measured with Sample Question Reference

Measures of intrusive thought content:
OII Thoughts related to contamination, harm to self Purdon and Clark
and others, and taboo behavior, including frequen- (1993)
cy and believability: rate frequency of “exposing
myself” from 0 “never” to 4 “always”
RRS-SF Continuous thoughts typical of depression sepa- Treynor et al.
rate from depressive symptoms, subscales: (2003)
Brooding—comparison between one’s current
state and some unachieved standard: What am I
doing to deserve this?
Reflection—repetitive thoughts about problem-
solving that might ameliorate negative affect:
Analyze recent events to understand why you are
depressed
Measures of the ability to control thoughts:
WBSI Difficulty in controlling thoughts, subscales: Wegner and
Intrusions—the degree of intrusion experiences: Zanakos (1994);
I have thoughts that I cannot stop. 2 factor structure,
Suppression—tendency/ability to rely on thought Schmidt et al.
suppression as a strategy: There are things I try (2009)
not to think about.
PSWQ The degree to which a person worries: Once I Meyer et al.
start worrying, I cannot stop. (1990)
TCQ Assesses strategies used to control thoughts: Wells and Davies
I think about something else. (1994)
IES-R Measures intrusions after stressful or traumatic Weiss (1997)
events: I thought about it when I didn’t mean to or
pictures about it popped into my mind.
including self-relevance, frequency, importance, orientation with regard to

goals, orientation with regard to social factors, and level of detail. Using a
hierarchical clustering analysis, four major dimensions of thought content
emerged: (a) level of construal (degree of temporal and perceptual specificity),
(b) degree of personal significance, (c) temporal orientation (future oriented
vs. past oriented), and (d) valence (positive, negative). Of relevance to the cur-
rent discussion, scores on these dimensions are associated with characteristics
related to mental health. For instance, higher levels of thoughts characterized
by negative valence and high levels of personal significance are associated
with higher levels of depression (Andrews-Hanna et al. 2013).

Name Construct Measured with Sample Question Reference
Hybrid measures of both content and control:
ROII Includes additional measures of the degree and Purdon and Clark
manner to which obsessive thought can be con- (1994)
trolled: rate 0 (never) to 5 (always), e.g., Say stop
to myself
ITQ Frequency, degree of distress, and degree of diffi- Dougall et al.
culty controlling intrusive thought: How disturb- (1999)
ing are these thoughts for you? How difficult is
it for you to get rid of these disturbing thoughts
when they occur?
EIS Frequency, unpredictability, unwantedness, Salters-Pedneault
interference, and distress caused by the intrusive et al. (2009)
thoughts after analog trauma: How often have you
found yourself thinking to any degree about the
rape scene since seeing the film?
IITIS Examines thought content and strategies for con- RCIF (2007)
trol via a semi-structured interview, e.g., identify
unwanted religious or immoral intrusions
Abbreviations:
ESI: Experience of Intrusions Scale PSWQ: Penn State Worry Questionnaire
IES-R: Impact of Event Scale-Revised, ROII: Revised Obsessive Intrusions
intrusion subscale Inventory
IITIS: International Intrusive Thought RRS-SF: Ruminative Response Scale,
Interview Schedule short-form
ITQ: Intrusive Thoughts Questionnaire TCQ: Thought Control Questionnaire
OII: Obsessive Intrusions Inventory WBSI: White Bear Suppression Inventory
Advantages, Limitations, and Potential Extensions
These types of self-report measures have advantages: they are typically short,
easily administered, generally well normed, can be used with both clinical and
nonclinical populations, and scores derived from them can be used as covari-
ates in adjunct analyses. In terms of limitations, they require metacognitive
abilities related to self-awareness and self-evaluation on the part of the respon-
dent, which may be compromised in individuals with more severe psychiatric
disorders. Perhaps most glaringly, however, is the fact that they can be narrow
in scope, as most were designed to address a specific psychiatric disorder. As
such, they tend to examine the types of processes (e.g., punctate vs. continu-
ous) and specific topics of intrusive thoughts that characterize a given psychi-
atric disorder.
Hence, a questionnaire on intrusive thoughts and their control is needed
that could be used more generally across individuals with a variety of clinical
disorders, as well as with individuals who do not meet clinical criteria. There
84 M. T. Banich
are numerous ways to design such a questionnaire, and it might be useful to

include the following capabilities:
1. Analyze thought content via the assessment of underlying common
factors across psychiatric and nonpsychiatric populations, such as their
valence and temporal orientation.
2. Assess the degree to which an individual has difficulty in controlling
thoughts as well as those individual differences that protect against in-
trusive thoughts (e.g., mindfulness) and the mechanisms by which they
act (e.g., Emerson et al. 2017).
3. Distinguish intrusive thoughts from mind wandering and task-unre-
lated thoughts (e.g., Maillet and Schacter 2016).
4. Contain optional subscales that could assess intrusive thoughts specific
to a given disorder (e.g., thoughts of contamination in individuals with
OCD) as well as differentiate those from thoughts that occur in other
psychiatric disorders.
Not only would the creation of such a questionnaire enable a finer assessment
of the nature of intrusive thoughts and their control, it might also enable as-
pects of intrusive thoughts to be linked to specific symptom clusters across
psychiatric disorders (e.g., fear).
Next, we take a look at other behavioral and experimental methods for ex-
amining intrusive thoughts. They are divided roughly into two types: the first
engenders intrusive thoughts whereas the second examines how thoughts (or
memory retrieval) can be controlled.
Engendering Intrusive Thoughts
To examine intrusive thoughts in an experimental setting, one approach is to

actually engender them. The purpose of this method, often referred to as the
symptom provocation paradigm, is to provoke intrusive thoughts, often of a
traumatic nature (e.g., Brewin and Saunders 2001). Generally, individuals
identify specific traumatic events in their lives or categories of stimuli that
engender intrusions in a pretest session. Then, within an experimental session,
specific pictures or stimuli that are traumatic are shown or individuals are ex-
posed to a category of stimuli associated with trauma, such as combat noise or
pictures for veterans (e.g., Daniels and Vermetten 2016).
A related approach, often referred to as analog trauma, is designed to in-
duce and engender trauma-related responses (e.g., intrusive memories). This
procedure typically involves having individuals watch a film that contains
graphic depictions of traumatic events, such as physical or sexual violence (for
a review, see Holmes and Bourne 2008). In an extension of this approach, vir-
tual reality can be used to induce an analog trauma (Dibbets 2019). Across both
symptom provocation and analog trauma, the degree and nature of thought
intrusions can then be examined within the context of the laboratory or via
a diary of intrusions for some specified time (e.g., one week) after exposure.
As discussed by Visser et al. (2018), a variety of points in memory pro-
cessing could be disrupted by or associated with intrusive thoughts: from the
original attention to and encoding of information to the access and retrieval of
memories. Procedures designed to engender intrusive thoughts can be com-
bined (a) to analyze the effects of other variables or manipulations at distinct
time points (before, during, or after exposure) and (b) to examine how differ-
ent processes (e.g., attention/focus, encoding, and recall) influence intrusions.
Manipulations implemented before exposure include having individuals recall
memories that induce high levels of self-efficacy (Krans et al. 2018) to put the
focus on themselves, or playing a distracting video game to place the focus
on something else (James et al. 2016b). Manipulations implemented during
viewing of the material include varying the cognitive load during the task or
instructions that lead to hyperarousal via hyperventilation (Nixon et al. 2007).
Manipulations after exposure have included varying instructions on how to
deal with the intrusions, such as rumination (“How can I drive again without
thinking about what could happen?”), integration so as to distinguish the video
from the person’s own nontraumatic experiences (“Think about your own driv-
ing experiences”), or distraction (“Try to recall as many African countries as
you can think of”) (e.g., Zetsche et al. 2009; Horsch et al. 2017).
The advantage of the symptom provocation and analog trauma methods is

that they can be implemented relatively easily. They also have face validity,
especially for syndromes such as PTSD, which is generally characterized by
intrusions linked to a specific event or trauma. These methods can also be
used with provocation for particular classes of items that might engender in-
trusive thoughts, such as those associated with OCD. Although individuals are
typically asked to report intrusions, there are other possible means of assessing
intrusions. For example, virtual reality approaches have the ability to provide
additional information: upon reimmersion into the environment, one can ex-
amine the degree to which an individual avoids that portion of the virtual real-
ity space associated with the traumatic scene (e.g., Dibbets 2019). With regard
to still photos, one could redisplay portions of a traumatic scene (e.g., a car
next to an overpass with a concrete barrier) without the specifically traumatic
context (e.g., a bloody person lying in the road by the car) and determine the
nature and duration of eye movements to the location of the trauma content.
Relatedly, eye movements might be employed to determine when individuals
are likely to be more inwardly focused, as has been used for lapses of attention
(or mind wandering) during reading (e.g., Reichle et al. 2010). Finally, diary
methods for recording intrusions after exposure could be expanded to use digi-
tal queries at random times via mobile apps and other smartphone technologies.
86 M. T. Banich
The vast majority of studies that examine intrusive thoughts do so from the
perspective of long-term memory formation and retrieval, which is particularly
appropriate when intrusive thought is driven by a specific event or circumstance,
such as occurs in PTSD. Far less work has focused, however, on mechanisms
related to intrusive thoughts, especially those not linked to a specific event, in
terms of how they get “stuck” in working memory and current consciousness.
Symptom provocation and analog trauma are not well suited to examining the
nature of recurrent, ruminative, and cognitively based intrusive patterns of think-
ing, which are typical in depression or worry (anxious apprehension) but cannot
be specifically linked to a particular point in time nor to a particular set of pro-
voking stimuli. For example, depressive intrusive thoughts often focus on how
one could “solve” the issues that lead to distress and negative affect. From think-
ing about social interactions with others to self-reflection on actions taken to an
analysis of one’s internal mood states, the topical range tends to be larger than,
for example, thoughts of contamination in OCD.
Engaging and Examining Thought Suppression Mechanisms
Another method of examining intrusive thoughts is to determine the effects

of formally trying to suppress an intrusive thought. In a classic version of this
task, individuals are given a period of time (e.g., five minutes) where they
are allowed to think of anything that comes to mind (often referred to as the
free-thinking condition). Afterward, they are placed in either an expression
condition, in which participants are told that they should think about a specific
item (e.g., a white bear) for a given period of time (e.g., five minutes), or in a
suppression condition, in which they are told to suppress thinking about that
item (i.e., the white bear) for an equal amount of time. The participant then
indicates by some means, such as ringing a bell or pressing a button, whenever
the item comes to mind (Wegner et al. 1987). This results in a paradoxical
effect: trying to suppress a thought at first leads to greater subsequent expres-
sion than if the idea had initially been expressed and then later suppressed
(Wenzlaff and Wegner 2000). Meta-analyses find a small to medium effect
of the rebound of thoughts after suppression in both clinical and nonclinical
groups (Abramowitz et al. 2001).
From this initial approach, several variations have been employed. In one
extension, individuals are asked to identify thoughts, images, or impulses that
pop into their mind unexpectedly and in an intrusive manner. The number of
intrusions during this free-thinking condition can be compared to a suppres-
sion condition as well as other potential manipulations, such as distraction in-
volving thinking about something else (e.g., a past or future weekend with
friends) or accepting a thought (e.g., think about the intrusive thoughts coming
out of your ears on little signs held by soldiers, who walk them in front and
then away from you) (e.g., Najmi et al. 2009).
Such measures have been used to examine individual differences in clinical

symptoms, cognitive abilities, or age. Researchers have examined, for exam-
ple, whether higher or lower levels of executive function and cognitive con-
trol influence the ability to suppress intrusive thoughts successfully in specific
populations, such as those with PTSD (Bomyea and Lang 2016), or whether
the suppression versus the expression of thoughts is linked to working mem-
ory ability (Brewin and Smart 2005). Using a different approach, others have
examined whether intrusions vary according to the content (more specific to
certain life periods) and age of the participants. In younger individuals, for
example, career success or failure might constitute the focus, whereas for older
individuals, memory loss (in particular, fear of forgetting friends and family)
might be more prevalent (Beadel et al. 2013).
Another experimental method that has been used widely in the laboratory to
examine control over thoughts, specifically memory retrieval, is the Think/No-
Think paradigm (Anderson and Green 2001). In this classic paradigm, indi-
viduals are taught associations between a cue word (e.g., “ordeal”) and a target
word (e.g., “roach”) to a given level of accuracy (e.g., 95%) that will ensure
a solid memory trace. In the experimental phase, some cues are presented so
that the participant must think about the associated target (Think condition),
while other cues are presented so that the participant should not think about
or allow the associated target into consciousness. Each cue is shown multiple
times so that there are numerous opportunities to exert cognitive control over
the memory of the target. Then in the test phase, the individual is shown cues
for each of the initial pairs, and memory for the associated target is assessed.
Memory is typically increased for Think trials and decreased for No-Think
trials relative to a baseline of items whose cues were not shown during the
experimental phase (which provides an index of forgetting since initial train-
ing). Hence, this paradigm is well suited to examine control over retrieval of
information from long-term memory, which is highly relevant to disorders in
which there is intrusive memory retrieval. Although initial studies used verbal
stimuli, similar effects have been observed for visual and emotional stimuli,
such as face–scene pairs (Depue et al. 2006). This may be more suited for
studying populations where intrusive thoughts take the form of images (e.g.,
object–scene pairs), such as in PTSD (Catarino et al. 2015).

These approaches provide methods for examining control over intrusive
thoughts as well as the mechanisms (e.g., suppression, distraction) by which
such control may be exerted. However, they rely on participant self-report of
the occurrence of those intrusive thoughts and a certain amount of metacogni-
tive awareness (i.e., internal monitoring of when those thoughts have occurred).
Although not specific to intrusive thoughts, the Think/No-Think paradigm pro-
vides a robust and tractable experimental paradigm to examine the control over
88 M. T. Banich
thoughts, especially with regard to memory retrieval. In clinical populations,

its use is limited due to the length of the procedure (e.g., 30 minutes to 1 hour)
and requirement that individuals learn and retain the pairs. Participants must
be able to perform the initial learning and sustain an adequate level of attention
and motivation to perform the task. As with the symptom provocation and ana-
log trauma approaches, these methods have clearer linkages to disorders like
PTSD and OCD than to the recurrent intrusive thoughts that are characteristic
of depression and anxiety.
Neural Systems Related to Intrusive Thoughts

A substantial body of research has focused on examining neural processes that
are associated with intrusive thoughts. In general, the main techniques used are
magnetic resonance imaging (MRI), typically used to localize brain systems
involved with intrusive thoughts, or electroencephalography (EEG), including
event-related potentials (ERPs) which provide information about the timing of
processes associated with intrusive thoughts. In general, these methods tend to
be used in combination with one of the behavioral approaches discussed above.
The utility of such approaches is that they can provide insight into the
mechanisms that may be generating intrusive thoughts. For example, although
one must be cautious in making reverse inferences from patterns of brain ac-
tivation (Poldrack 2011), evidence of prefrontal activity when attempting to
limit the intrusiveness of thoughts is suggestive of an active control process,
whereas evidence of activity in subcortical regions, such as the basal ganglia,
would be suggestive of a more automatized process. Likewise, alterations in
early ERP components (e.g., P1, N1) are more suggestive of attempts at control
over sensory aspects of an intrusive thought, where alterations in later ERPs
(e.g., P3 and N4) would be more suggestive of control over information in
working memory or of a semantic nature, respectively.
Brain Processes Associated with Intrusive Thoughts

Measures Used to Engender Intrusive Thoughts
The symptom provocation paradigm, especially as it relates to individuals with

PTSD, has been migrated into a neuroimaging environment. As with all func-
tional MRI (fMRI) studies, the condition of interest must be contrasted with a
baseline of some sort that does not engage the behavioral construct of interest.
Often in these studies, brain activation in a symptom provocation condition
is compared to baseline condition; this may involve processing information
from a nontraumatic memory, emotionally neutral pictures (e.g., civilian or
noncombat scenes), or non-emotional information (e.g., white noise or rest).
Meta-analyses across such studies of individuals with PTSD (e.g., Sartory et
al. 2013) have found that these paradigms reliably isolate a set of brain regions
that differentially activate during the symptom provocation as compared to

comparison conditions. This set of regions includes portions of the default
mode network, considered to be involved in internal thought, self-referential
processing, and autobiographical memory (Andrews-Hanna et al. 2014), as
well as areas that process the emotional significance and valence of informa-
tion, such as pregenual portions of the anterior cingulate and the amygdala.
The involvement of these regions in autobiographical memory and emotion
processing provide a piece of converging evidence, not provided by behavioral
paradigms alone, that the symptom provocation technique is effective at induc-
ing reexperiencing.
With regard to EEG/ERP methods, some studies (Roh et al. 2017) have ex-
amined differences in specific ERP components (e.g., error-related negativity)
under symptom provocation as compared to other conditions in individuals
with disorders characterized by intrusive thoughts (e.g., in individuals with
OCD). Although relatively rare, other studies have examined EEG metrics,
such as the hemispheric asymmetry of frontal alpha rhythms (as an index of
approach and avoidance behaviors), to symptom provocation (for a review, see
Meyer et al. 2015).
Measures Used to Examine the Experience of and Control over

Intrusive Thoughts
One can also utilize neuroimaging techniques, in conjunction with behavioral

methods that index when an intrusive thought occurs or when control sys-
tems are engaged, to limit or otherwise attempt to suppress such thoughts.
One approach is to measure brain activation during the time periods in which
intrusions occur and compare that to activation during time periods without in-
trusions. In some paradigms, the participant notes in real time when the intru-
sion occurs by pressing a button. Brain activation during the intrusive thought
is then compared to some baseline, such as the time period right afterward
when re-suppression occurs (e.g., Carew et al. 2013). Similarly, in the Think/
No-Think task, one can examine neural activation on unsuccessful No-Think
trials in which the item to be suppressed intrudes upon consciousness. This
activation can be compared to Think trials, in which controlled (rather than
intruded) retrieval has occurred (Hellerstedt et al. 2016), or to No-Think tri-
als, in which the item is successfully suppressed (Levy and Anderson 2012).
Examining brain activation during intrusions has been done with fMRI, al-
though the fine temporal resolution of EEG/ERPs may be better suited. For
example, EEG/ERPs can be used to provide a putative index of how long the
intrusion remains in working memory (Hellerstedt et al. 2016) or to identify
the onset of the process that is engaged to keep it from coming into working
memory (Castiglione et al. 2019).
While some studies look at real-time intrusions, which provide a “state”
perspective, other approaches examine this issue from a “trait” perspective.
90 M. T. Banich
Here, individuals are characterized as to the degree to which intrusions are

experienced during their daily lives or over longer time periods. For instance,
using daily diary entries, Kuhn et al. (2013) examined the average degree to
which intrusive thoughts are experienced by an individual over a six-month
period and then linked the intrusion rate to patterns of brain activation at rest.
Rather than focusing on intrusions in particular, another approach com-
pares brain activation assessed by fMRI across various conditions, such as free
thought, suppression of a given thought, or suppression of all thoughts (e.g.,
Wyland et al. 2003). Another way is to examine mechanisms that are involved
in suppressing thoughts versus replacing thought (e.g., Benoit and Anderson
2012). As detailed below for aspects of working memory, mechanisms of item
replacement, specific item suppression, and suppression of all thoughts appear
to have partially overlapping but distinct neural mechanisms (Banich et al.
2015). These findings support separate consideration of these potential mecha-
nisms of controlling intrusions.
Another issue that can be fruitfully examined using neural investigations
is the degree to which the processes involved in suppression of thoughts are
similar to or distinct from other categories of suppression. While a backward
inference from brain activation to cognitive processes must be performed
with caution, neuroimaging studies nonetheless can provide insights into the
specificity of control over memory versus other processes. For example, in
the same individuals, Depue et al. (2016) examined the degree to which acti-
vation during a memory suppression task (measured by the Think/No-Think
task) engendered similar or separate neural mechanisms than either the sup-
pression of emotion or the suppression of motoric responses (all compared to a
domain-appropriate nonsuppression baseline). While all three tasks produced
activation in right dorsolateral prefrontal cortex, it was the connectivity of this
region to domain-specific processing regions (e.g., the amygdala in the case of
emotion regulation) that differentiated these three types of suppression. Other
studies suggest somewhat overlapping mechanisms of memory and emotional
suppression (Gagnepain et al. 2017) as well as memory and motoric suppres-
sion (Castiglione et al. 2019).
The advantage of capturing brain processes associated with both the engender-
ing and controlling of intrusive thoughts is that they can provide more informa-
tion than a simple behavioral reaction time (i.e., button press) or retrospective
report. One must be cautious in inferring the engagement of cognitive pro-
cesses from patterns of brain activation, even within the context of the broad
set of knowledge regarding the neural circuitry underlying memory processes.
Nonetheless, these patterns provide insight into what aspects of memory pro-
cessing are disrupted during intrusive disorders, or whether control mecha-
nisms are intact but mainly engaged at inappropriate times. A disadvantage of
such approaches is that these neural metrics often require multiple trials for
signal averaging; thus, the frequency of intrusions poses a potential limitation.
Although they may readily occur, with increasing practice at suppression, they
tend to become less frequent (Hellerstedt et al. 2016), which may limit the
amount of data that can be collected.
The Multiplicity of Brain Metrics Available
The studies discussed above that used fMRI focused primarily on brain re-
gions that become active during an intrusion or during the attempt to control
an intrusion. Additional metrics, however, should be examined to see whether
they can provide a distinct window into these processes. For instance, activa-
tion within cognitive control (e.g., dorsolateral prefrontal cortex) and mem-
ory-related regions (e.g., the hippocampus) has been implicated in suppressing
memory retrieval, as has the connectivity between these regions (e.g., Depue
et al. 2007; Benoit et al. 2015). Connectivity patterns could be examined us-
ing independent component analysis, which reveals groupings of brain regions
whose activity follows a similar temporal time course during the suppression
of a thought, as compared to other processes, such as visual imagery (Aso et
al. 2016).
A variety of electromagnetic techniques can be applied to studying intrusive
thoughts. These may focus on specific ERP components, such as the parietal
old/new component, which occurs approximately 50–80 ms after stimulus pre-
sentation and is thought to be an index of memory retrieval (Rugg and Curran
2007). Such measures could be combined with measures of neural oscillations,
recorded from the scalp (e.g., Depue et al. 2013) or via intracranial record-
ings in patients undergoing surgery for epilepsy (Oehrn et al. 2018), to pro-
vide information on the control of such retrieval. Magnetoencephaology has
been used to examine downregulation of sensory aspects of long-term memory
in the gamma band (70–120 Hz) in traumatized refugees (Waldhauser et al.
2018). Optical imaging methods (e.g., functional near infrared spectroscopy,
which provides information on both the location and time course of activation)
have been used to examine brain activation in individuals with PTSD during
symptom provocation (Gramlich et al. 2017) as well as in individuals high in
rumination during stress (Rosenbaum et al. 2018).
All of these methods record or otherwise observe the nature of brain ac-
tivation associated with memory retrieval or control processes. In contrast,
current work that focuses on using brain stimulation techniques (e.g., transcra-
nial magnetic stimulation or transcranial direct current stimulation) to alter
intrusive thoughts (i.e., to induce or disrupt them) is still preliminary. In one
study, brain stimulation of prefrontal regions and the underlying white matter
in three patients about to undergo surgery for epilepsy was found to induce
intrusive thoughts. For example, when stimulated one patient reported: “The
stimulation induces the disappearance of the word in my mind and replaces
92 M. T. Banich
it with something else” (Popa et al. 2016:3). Another reported that he had “a
thought that seems to come from nowhere” (Popa et al. 2016:4). To the best
of my knowledge, brain stimulation techniques have not been used to disrupt
thought. In addition, such methods may provide other insights into intrusive
thoughts. For example, aspects of intrusive memories in PTSD tend to be over-
generalized (Brewin 2011); that is, memories are not clearly differentiated.
Transcranial direct current stimulation over lateral occipital cortex during the
encoding of a memory leads to interference between memory representations,
presumably because of coactivation and less differentiation between those rep-
resentations (Koolschijn et al. 2019). Hence, such stimulation might poten-
tially be used as a system to model aspects of intrusive thoughts in PTSD.
Brain-based methods offer a wide variety of tools and a number of different

metrics (e.g., brain activity, brain connectivity) that can be used to explore
the mechanisms that underlie intrusive thinking. They provide converging
evidence for purported mechanisms of intrusive thought and can be used to
distinguish potential mechanisms involved in memory control and retrieval of
intrusive memories. In addition, brain-based measures offer unique insights.
For example, brain-imaging techniques have indicated that memory retrieval
can be actively suppressed, as evidenced by a reduction below baseline in
activation of hippocampal regions during attempts not to think about specific
items (Depue et al. 2007). Recent advances in brain-imaging techniques allow
information about intrusive thoughts to be gleaned from nonstructured and
more naturalistic stimuli (e.g., a movie) without requiring a specific contrast
between conditions (Huk et al. 2018). This opens the possibility for sophisti-
cated computational algorithms to extract over time those critical patterns or
signatures of brain activity that are associated with the formation or retrieval
of intrusive thoughts.
Individual Differences: Approaches to Brain Anatomy and Function

Associated with Intrusive Thought
Another approach is to examine how the functioning of neural systems varies,

depending on differences among individuals in the degree of intrusive thoughts
and/or the degree to which they can control such thoughts. Some studies ex-
amine aspects of the brain that are relatively static (e.g., brain anatomy or the
organization of intrinsic resting-state networks) in individuals who experience
high levels of intrusive thoughts: PTSD or OCD patients (e.g., Chen et al.
2018a; Gürsel et al. 2018) or individual self-reports from single or extended
time periods (e.g., Kühn et al. 2013). While such studies may provide informa-
tion about variation in potential brain structures involved in intrusive thought
(e.g., the hippocampus), they may not provide information specific to intrusive
thinking. For example, a stress response associated with traumatic events (e.g.,
increased cortisol and neurotoxicity of hippocampal cells) could cause reduc-
tion in hippocampal volume or shape. Moreover, examining brain anatomy
and resting-state connectivity may not be ideal for studying intrusive thought,
because both are relatively static, whereas intrusive thoughts, by nature, are
time limited and dynamic.
Other approaches characterize individuals according to their level and/or
controllability of intrusive thought to determine how these factors might af-
fect brain activation. For example, during suppression of recently experienced
items, individuals with a higher degree of self-reported difficulty in removing
current thoughts from consciousness had higher levels of activation in Broca
area; this presumably represents an inclination toward inner speech (Banich
et al. 2015). Another recent and potentially profitable approach is to use mag-
netic resonance spectroscopy (MRS) to examine potential neurochemical
mechanisms that may enable certain individuals to block memory retrieval.
For example, individuals with higher levels of GABA in the hippocampus, as
assessed by MRS, have a greater ability to suppress memory retrieval (e.g.,
Schmitz et al. 2017). The disadvantage of using MRS methods is that they are
quite time consuming (e.g., 25 minutes). In addition, only a few brain regions
can be interrogated during a scanning session, and the brain region interro-
gated is much larger (e.g., at least 3–4 times greater) relative to functional
neuroimaging methods.
Exploring individual differences in the psychological processes involved in in-

trusive thought has a long and fruitful history, and can be equally well applied
for use with neural markers. However, for neuroimaging, an individual differ-
ences approach generally requires a larger sample size to detect covariation
than is required in studies designed to detect group average patterns of activa-
tion (Cremers et al. 2017). Thus, utilizing an individual differences approach
requires more time and money.
Future Frontiers for Brain-Imaging Methods
Using Brain-Imaging Methods in a Predictive Manner
There is much interest in determining whether an individual will experience

intrusive memories after a distressing event. Prior work has examined whether
certain characteristics of an individual and/or the way in which a distress-
ing event is processed are robust predictors of subsequent intrusive thoughts
(e.g., Marks et al. 2018). This work has been extended to examine whether
measures derived from brain metrics might predict subsequent intrusions. For
94 M. T. Banich
example, ERP measures of the effectiveness of suppression (greater amplitude

of a fronto-centrally distributed N2) from the Think/No-Think task can predict
subsequent intrusions after analog trauma (Streb et al. 2016). Relatedly, pat-
terns of brain activation derived from machine-learning techniques, during the
encoding of material in an analog trauma paradigm, can be used to predict the
degree to which an individual will have subsequent intrusive thoughts (Clark et
al. 2014b). Although in its infancy, such approaches may have much potential.
Neural Markers versus Self-Report
One important limitation of many of the methods described above is that they
rely on self-report to verify an intrusive memory or control over thoughts, and
thus do not provide insight into the nature of the representation of that memory.
Neurally based measures have been used to try to address this issue.
Autonomic Measures
Autonomic measures have been used in conjunction with the behavioral

methods discussed above. The idea is that reexperiencing traumatic events
should induce physiological changes and that successful suppression of
such thoughts should be associated with reductions in such physiological
responses (e.g., May and Johnson 1973). Such measures have also been used
in conjunction with an individual differences approach. For example, greater
heart rate variability is associated with a better ability to inhibit thoughts,
either in a structured thought suppression situation, the Think/No-Think
paradigm (Gillie et al. 2014), or in the self-report of intrusive thoughts over
specific periods of time (Gillie et al. 2015). Such physiological measures,
however, are mainly nonspecific in nature and could reflect arousal, emo-
tional distress, or anxiety, either in a state or trait manner that is unrelated to
intrusive thoughts.
Neuroimaging Approaches
Although much work on intrusive thinking has focused on the retrieval of in-
formation from long-term memory, by nature, intrusive thoughts involve ac-
cess to and active representation in working memory (for further discussion,
see Visser et al., this volume). Understanding whether a thought is currently
in the focus of attention in working memory is an important issue. Initial work
suggests that brain-imaging techniques can be applied quite fruitfully to verify
that individuals are indeed experiencing specific thoughts and/or manipulat-
ing them. In one study (Banich et al. 2015), individuals were shown a picture
or heard a brief snippet of a familiar tune (e.g., “Happy Birthday”) for four
seconds; immediately afterward they then had to manipulate the item in one of
four manners: maintain it, replace it with a different image/tune, specifically
suppress the item, or clear their mind of all thought. Providing at least some ev-
idence that participants were indeed manipulating their thoughts as instructed,
a significant reduction of activity in primary sensory areas (e.g., visual cortex,
auditory cortex) was observed averaged across all trials for the two conditions
in which a thought needed to be cleared and removed (suppress the item or
clear their mind of all thought), compared to when there was an active thought
(maintain or replace). In addition, results indicated that while some neural
mechanisms were commonly engaged across these operations (e.g., across re-
placing, suppressing, or clearing an item as compared to maintaining it), there
are also specific neural mechanisms that differentiate the suppression of an
item from clearing one’s mind of all thought (Banich et al. 2015). Thus, neuro-
imaging may provide a neural marker and confirm specific mental operations
performed on a given thought. This may lead us to differentiate the ways in
which thoughts can be removed from working memory.
In follow-up work, machine learning was incorporated to expand the ques-
tions and issues that can be examined (Kim et al., submitted). Specifically,
via a localizer task, a machine-learning classifier was used to distinguish spe-
cific categories of items (e.g., places, faces, fruit). These classifiers were then
used on a trial-by-trial basis to determine the degree to which removing the
thought was successful. If a thought is successfully removed, then the clas-
sifier fit should be poor. If the thought is maintained, then the classifier fit
should be high. This approach enables us to examine the nature of the mental
representation on a trial-by-trial basis and provides a means to determine the
level at which such representations are maintained and/or cleared via specific
category and subcategory classifiers (e.g., fruit: apples, grapes, pears). For ex-
ample, one can examine whether clearing the thought of an apple generalizes
to other fruit (e.g., grapes and pears). In addition, patterns of brain activation
can be examined as a function of classifier fit to determine which brain re-
gions are highly active when the classifier fit is low (indicative of clearing the
thought), compared to when the classifier fit is high (indicative of the thought
remaining). This, in turn, may provide insight into which brain regions are
involved in exerting control over thoughts. While this study did not explicitly
track intrusions, such methods could be extended to track the content of intru-
sive thoughts on a real-time basis by identifying multivariate neural patterns
of distinct forms of intrusive content and evaluating the degree to which these
patterns manifest on a moment-to-moment basis.
In summary, there are a variety of interesting new directions that might be
used to provide novel insights into the nature of intrusive thought. These range
from adapting and using paradigms from other areas of cognitive and affec-
tive psychology to investigate intrusive thinking, to new approaches and ap-
plications from brain-imaging methods that might be used to verify or predict
intrusive thought.
7
Conceptualizing Intrusive
Thinking at the Level of
Psychological Mechanisms
Marie-Hélène Monfils and David M. Buss
Abstract
Intrusive thinking, the sudden occurrence of unwanted thoughts, images, or impulses, is

a frequent and natural occurrence within our stream of consciousness (Clark and Purdon
1995). Present in both clinical and nonclinical samples, the high incidence of intrusive
thoughts across the population renders challenging the task to identify meaning behind
their occurrence. Their presence, frequency, and content do not appear, however, to be
random. Intrusive thinking manifests differently in clinical versus nonclinical popula-
tions. They may be associated with certain emotions, thus offering a glimpse into their
potential adaptive nature. This chapter examines what intrusive thoughts are and what
they are not. It explores how they manifest differently in clinical versus nonclinical pop-
ulations and asks whether these different presentations can provide insights into their
origin. It evaluates intrusions as possible manifestations of adaptations and examines
intrusions linked to evolved emotions (e.g., fear, rage, jealousy, and love). Identifying
the possible reasons behind intrusive thinking may help guide future treatment.
Introduction
A commuter experiences a sudden urge to jump off the subway platform as the

train arrives at the station. An individual engaged in cleaning up after dinner
suddenly has a vivid image of throwing a plate against the wall, as a fit of rage
intrudes their thoughts. Someone looking through their wardrobe to find some-
thing to wear that day suddenly hears a voice in their head saying, “You’re
such a loser.” As another person walks their dog, a sudden image of repeatedly
stabbing a passerby jarringly interrupts their train of thought. Another is driv-
ing to work when a voice interrupts their thoughts and suddenly proclaims:
“You need to get on stage with a guitar! You’re a rock star!”
98 M.-H. Monfils and D. M. Buss
Intrusive thinking, the sudden occurrence of unwanted thoughts, images, or

impulses, is a frequent and natural occurrence within our stream of conscious-
ness (Clark and Purdon 1995). Intrusions occur in both clinical and nonclinical
samples, and their high incidence (80–90%) across the population renders chal-
lenging the task to identify a meaning behind their occurrence (Clark 2005).
It does appear, however, that their presence might not simply be random. The
manifestation of intrusions differs in clinical versus nonclinical populations,
and they may be associated with the presence of certain emotions that offer
glimpses into their potential adaptive nature.
Here, we examine what intrusive thoughts are and what they are not. We
explore how intrusive thoughts manifest differently in clinical versus nonclin-
ical populations. We ask whether the different presentations in clinical and
nonclinical populations might provide insights into their origin, and evaluate
intrusions as adaptations.
What Are Intrusions?
According to Clark (2005:4), an intrusive thought is “any distinct, identifiable

cognitive event that is unwanted, unintended, and recurrent. It interrupts the
flow of thought, interferes in task performance, is associated with negative
affect, and is difficult to control.” This definition is generally consistent with
others used to describe the phenomenon (Beck 1967; Horowitz 1975; Klinger
1978; Rachman 1981) and appears to include the following characteristics
(Clark 2005):
• It is a distinct thought, image, or impulse that enters conscious
awareness.
• It is attributed to an internal origin.
• It is considered unacceptable or unwanted.
• It interferes with ongoing activity.
• It is unintended.
• It tends to be recurrent.
• It easily captures attentional resources.
• It is difficult to control.
Clark also suggests that intrusive thoughts are negative, although it is unclear
whether this criterion is universally accepted, a point to which we return below
(e.g., Gregory et al. 2010).
Images versus Thoughts
Intrusive thoughts are acknowledged to manifest in different ways. Sometimes

they present as images or scenarios, and other times as an internal voice devoid
of imagery.
Psychological Mechanisms 99
Memories versus Nonmemories
Though there is evidence that intrusive thoughts may be born from the recall
of previous experiences, this is not always the case. Intrusions can sometimes
consist of memories and other times not. There are instances in which the in-
trusion is autobiographical, but it is uncertain if that is always the case.
Spontaneous versus Triggered
Some intrusive thoughts are triggered by a cue. For example, an individual

may experience an intrusive thought compelling them to stab themselves in
the chest upon seeing a kitchen knife on the table. Seeing the knife might have
served as a cue that prompted the intrusion. In other cases, the intrusion may
seem to appear out of nowhere, without an obvious cue having acted as a trig-
ger. The fact that a triggering cue is not identified does not necessarily mean
that it was not present. Rather, it could be that the cue was subtler and not
explicitly observed.
Valence of Intrusive Thoughts
Many definitions of intrusive thoughts imply that they are negative in valence
(Clark 2005); however, Gregory et al. (2010) propose that intrusions may actu-
ally present as highly positive in individuals experiencing a hypomanic state.
These could be similar in content to thoughts related to delusions of grandeur.
Other situations suggest that positive intrusions exist, including in nonclinical
samples. One example is that of infatuation in which an individual experiences
intrusive thoughts about a loved one, and many such intrusions carry a posi-
tive valence (e.g., fantasies of union or sexual consummation). Even situations
that appear to have a negative connotation on the surface could carry positive
valence for the individual experiencing the intrusion. For example, imagining
the suffering of an enemy could be quite positive in a scenario of homicidal
ideation.
What Predicts Intrusive Thoughts?
Intrusive memories can be triggered by rumination, a phenomenon that is often

present in individuals suffering from anxiety, depression, or both (Birrer et al.
2007). A number of mental health disorders are associated with the presence
of intrusive thoughts (discussed further below); however, they also manifest
in nonclinical samples. The overall incidence appears quite high: 80–90% of
individuals in nonclinical samples report experiencing intrusions (Clark 2005).
Below, we discuss possible origins of intrusions. There is evidence suggesting
that attaching meaning or importance to intrusions can impact their frequency
and controllability (Freeston et al. 1991).
What Intrusive Thoughts Predict
Intrusive thoughts are quite prevalent following trauma, although findings

suggest that their frequency and severity is not predictive of posttraumatic
stress disorder (PTSD) symptomatology (McFarlane 1988; Shalev 1992).
Interestingly, Brewin et al. (1998) found that while the presence of intrusive
memories either at baseline or prior to follow-up made an additional significant
contribution to anxiety, it did not affect depression at follow-up (Brewin et
al. 1998). Having high-frequency involuntary intrusive memories at baseline,
however, significantly predicted later depression, even when controlling for
the severity of symptoms at baseline (Brewin et al. 1999).
One interesting feature of intrusive thoughts is that they are a common fea-
ture across multiple psychiatric disorders.
Intrusions in Clinical Populations
The incidence of intrusions is quite high across multiple mental health disor-
ders, where they are known to occur in individuals with obsessive-compulsive
disorder (OCD), generalized anxiety disorder (GAD), PTSD, body dysmor-
phic disorder, eating disorders, depression, bipolar disorder, and others. The
manifestation of intrusions appears to be partly affected by specific diagnoses.
For example, an individual with OCD who engages in extreme handwashing
might experience germ intrusions, whereas an individual with body dysmor-
phic disorder might get intrusions related to food items. For a more extensive
discussion of intrusions in clinical populations, see Schlagenhauf et al. and
Visser et al. (this volume).
Interestingly, intrusions appear to be dissociable from other characteristics
that might be more specific to only one or two disorders. Whereas obsessions
are thought to be characteristic of OCD, worry is a central feature of GAD
(although not exclusive to it), and negative thoughts and rumination may typi-
cally be present in individuals with depression; intrusions are often present in
all of these conditions. Let us now compare and contrast intrusions with worry,
rumination, obsessions, and negative thoughts.
Intrusions versus Worry
Defined as “a chain of thoughts and images, negatively affect laden and rela-
tively uncontrollable” (Borkovec et al. 1983), worry is a central feature in GAD,
but it also occurs with high incidence in nonclinical individuals. In definition
and in practice, worry and intrusive thinking are quite similar: They both inter-
rupt ongoing thoughts and activities, and they can both present as thoughts or
images, although worry occurs more frequently as verbal and intrusions more
frequently as images (Clark 2005). Intrusions are thought to be less voluntary
than worry; that is, worry can be brought on volitionally, whereas intrusions
are by definition involuntary and disruptive. Another distinctive feature that
disambiguates intrusions from worry is that intrusions are generally discrete
and brief, whereas worry need not be.
Intrusions versus Rumination
Defined by Nolen-Hoeksema and Morrow (1991) as repetitive and passive

thinking about one’s symptoms of depression, rumination is to depression
what worry is to GAD (Borkovec et al. 1998). To our knowledge, no study has
directly compared the differences between intrusions and rumination; how-
ever, respective reports for each provide clues as to what disambiguates the
two. Whereas intrusions are thought to be brief, sudden, and to involve gener-
ally unwanted thoughts or images, rumination involves a train of thought that
is longer, repetitive, and recurrent (Clark 2005). It is possible that intrusions
may trigger rumination, which in turn may precipitate a depressive or anxious
episode. As such, the same content may be at the source of intrusions and ru-
mination. In thinking about the distinction between intrusions and rumination,
one might imagine that an intrusion could occur during rumination.
Intrusions versus Obsessions
Obsessions and intrusive thoughts are very similar, where the former appears to
be an extreme version of the latter. Another characteristic that helps dissociate the
two is that intrusions may sometimes be irrelevant to the self, whereas obsessions
are relevant. Obsessions may often prompt behaviors such as compulsions that
are intended to diminish the associated thoughts and manifest as OCD.
Intrusions versus Negative Thoughts
Intrusions can be dissociated from general negative thoughts in that the former
is more likely to be irrational, whereas negative thoughts are more likely to be
rational. In this context, rational refers to thoughts that are not at odds with the
present context. An individual might be experiencing negative thoughts about
their promotion prospect during a recession, for example. If, during a positive
economy and after receiving a positive evaluation, they jarringly internally
hear the words “you’re about to get fired” just prior to giving an important
presentation, their experience was an intrusion. Intrusive thoughts are more
disruptive of day-to-day activity than general negative thoughts. Negative
thoughts are a core characteristic of individuals with depression, and generally
manifest as “thoughts” or in a verbal way rather than images. Intrusions may
present either as verbal or as images, most commonly the latter. Unlike other
forms of negative “processing,” intrusions seem to be relatively common in
nonclinical populations.
Manifestation of Intrusions in Nonclinical Populations

and the Origins of Intrusive Thoughts
Little is known about the etiology of intrusive thoughts. Different theories have
been proposed, but as yet we do not have a practical understanding of intru-
sions’ origins. Here, we briefly review various theories on intrusions, examine
their manifestation in nonclinical samples, and discuss the parallels between
intrusions and memory retrieval.
Theories on the Etiology of Intrusions
Salkovskis (1988) suggests that intrusive thoughts might be an inherent aspect

of problem solving. He proposes that despite being disruptive to thinking in
the moment, intrusions may be useful, and that the very reason they appear
suddenly and are intrusive and compelling could be that they are meant to
be noticed. In other words, if intrusions appeared as a simply nondisturbing
thought, we might not pay attention to them.
Rachman’s view on intrusive thoughts is predominantly based on the etiol-
ogy of obsessions (Rachman 1981). He also believes that an important con-
tributing factor to intrusions is the development of a mood state that sets the
tone for intrusions to occur. Rachman proposes, for instance, that individuals
who are stressed and in a dysphoric mood state are more likely to experience
intrusions. In such cases, individuals are also thought to have greater difficulty
ignoring or suppressing the intrusive thoughts. He also suggests that certain
personality characteristics (e.g., neuroticism, heightened anxiety) may make
individuals more susceptible to experiencing intrusions.
Klinger (1978) proposes that intrusive thoughts are associated with “current
concerns.” In other words, intrusions occur when thinking is interrupted and
the thought process shifts toward addressing what was brought about by the
intrusion (the current concern). The intrusions, then, can be external cues or
nonverbal events (Klinger 1999).
Horowitz proposed a reformulation of intrusive thoughts based on psycho-
analysis. In his account of intrusive thoughts, Horowitz posits that active mem-
ory storage is characterized by an intrinsic tendency to repeat its represented
contents, which continues until the storage of contents in active memory is ter-
minated. Appropriate cognitive processing of the memory content terminates
the process. Horowitz proposes that stressful events may yield intrusions that
stimulate an active memory of an experience. This memory activation occurs
repeatedly until there is integration of old and new information, perhaps to
reconcile representations of a memory with a person’s inner view of the world
(Horowitz and Wolfe 2003). Horowitz’s formulation appears particularly rel-
evant to traumatic memories.
The general overarching theme across the views held by Salkovskis,
Rachman, Klinger, and Horowitz is that, disruptive and disturbing as they are,
What?
Why? How?
When?
Jumping on train Experiencing depression Cognitive behavioral

tracks intrusion therapy, pharmacotherapy
Standing too close to the track Adjust one’s position
Figure 7.1 Identifying why intrusions arise may help guide how we can best treat them.
intrusions might actually serve an adaptive purpose. One difficulty in these

interpretations of intrusive thoughts is that they are not exactly practical. That
is, it is difficult to conceive how such theories might guide the development
of future treatment. We believe there is value in examining the manifestation
of intrusive thoughts in nonclinical populations and to extract the possible un-
derlying adaptive basis for their presence. In short, if we are poised to identify
why the brain produces intrusive thoughts, we should be better equipped to
determine how to address them (Figure 7.1). What problem is the brain trying
to solve? Are intrusive thoughts inherently harmful? Do they represent a ben-
eficial mechanism gone awry?
Informed by these theories, we present an adaptationist perspective on in-
trusive thoughts. Thereafter, we examine the manifestation of intrusions in
nonclinical samples through specific examples and extract two elements (con-
tent and process) that might provide useful insight into treatment avenues for
mental health disorders in which intrusions are often present.
An Adaptationist Perspective on Intrusive Thoughts
From the perspective of modern evolutionary biology, adaptations are char-

acteristics that evolved because they contributed in a specific way to solv-
ing a problem or challenge tributary to successful survival or reproduction
(Williams 1966). Propensities to become fearful of snakes and spiders, for
example, evolved because they led their bearers to avoid these dangers to sur-
vival (Öhman and Mineka 2001). Evidentiary criteria for invoking adaptation
include economy, efficiency, and, importantly, improbable precision of func-
tional design.
From this perspective, we ask: Do intrusive thoughts in nonclinical popula-
tions show evidence of functional design? Any sensible answer is reliant on
further conceptual and empirical work. Guided by an adaptationist perspective,
we offer a few preliminary suggestions or heuristics which rely on the follow-
ing metatheoretical premises:
1. Organisms have finite time and resource budgets.

2. Organisms have evolved decision rules to prioritize effort allocated
to some adaptive problems at the expense of others in temporal se-
quences: when faced simultaneously with tasty-looking ripe fruit, an
attractive mate, and a dangerous snake, for example, humans prioritize
effort allocated to avoiding a lethal snake bite, postponing effort de-
voted to the other adaptive problems.
3. Emotions such as fear, rage, disgust, and jealousy mobilize attention
and effort to specific threats or challenges, orchestrating an organism’s
cognition, physiology, and behavior to address those challenges (Tooby
and Cosmides 2008; Al-Shawaf et al. 2016).
One novel hypothesis that we are proposing here is that intrusive thinking may
be one important design feature of evolved emotions that have this prioritiza-
tion function, directing attention and allocating effort to solving some adaptive
challenges at the expense of others.
An important feature of these adaptations is their probabilistic nature,
guided by error management logic. Probabilistic nature simply means that ad-
aptations only succeed in solving adaptive challenges with some likelihood,
not invariantly. Although there is compelling evidence for evolved fears of
snakes and spiders (e.g., Öhman and Mineka 2001), and these adaptations have
undoubtedly saved many lives of their bearers, these adaptations do not invari-
ably prevent life-threatening bites: more than 81,000 people worldwide die
each year from snake bites. Evolved fears function probabilistically.
Error management theory is a metatheory of decision rules, combining
signal detection theory with evolutionary theory (Haselton and Buss 2000).
At an abstract level, when confronted with uncertain environments, there are
two possible ways to err inferentially: making false positives and making false
negatives. When recurrent cost asymmetries of making these two types of er-
rors exist over evolutionary time, selection will favor decision rules to avoid
the costlier error, even if they result statistically in more frequent errors. When
perceiving a rustle afoot in a thick grassy wooded area, for example, one can
err by inferring that a snake is absent when it actually is present (false nega-
tive) or by inferring that a snake is present when it is not (false alarm). In this
example, failing to detect an actual dangerous snake in the grass is a costlier
error than falsely inferring a snake’s existence when there is, in fact, no real
threat. Error management theory, in this case, predicts that fears of this sort
have evolved to avoid the costlier error, generating avoidance of probabilistic
threats, some or many of which will turn out to be false alarms. Although
these evolved systems are biased, they are adaptively biased. Error manage-
ment theory has garnered much empirical support, leading to the discovery of
phenomena ranging from the auditory looming bias and the vertical descent il-
lusion in the perceptual domain to the sexual overperception bias and infidelity
overinference bias in the social domain (Haselton and Nettle 2006).
A final element in this framework, as applied to intrusive thoughts, is the

mismatch principle, which states that evolved traits that were adaptive in the
ancestral environments in which they evolved may misfire and become mal-
adaptive in modern evolutionarily novel contexts (e.g., Spinella 2003; Li et al.
2018a). A prime example is eating disorders. Humans evolved in food-scarce
environments and have evolved feeding adaptations to consume calorie-rich
substances—those high in fat and sugar—when encountered and to easily store
metabolic surpluses in the form of fat deposition. In modern environments that
contain an abundance of these resources, easily obtainable with minimal effort
in concentrated forms (e.g., fast-food restaurants or grocery stores), humans
tend to overeat. Obesity and type 2 diabetes, absent in traditional hunter-gath-
erer cultures, are largely the result of these evolutionary mismatches, along
with other factors, such as more sedentary living.
The elements of these principles lead to the hypothesis that intrusive
thoughts are functional parts of evolved emotion systems. They are designed
(in part) to mobilize attention and effort toward specific adaptive challenges.
They are often adaptively biased, designed to avoid costly errors at the expense
of more frequent errors. Some are maladaptive in modern mismatched envi-
ronments that are widely discrepant from the ancestral environments in which
they evolved.
Possible Adaptive Purpose of Intrusions in Nonclinical Samples
Individuals must constantly make decisions over choices and prioritize task
importance. It is conceivable that intrusions act as a means to emphasize
what should be worked through as soon as possible, in line with Salkovskis,
Rachman, and Klinger. The fact that intrusions appear suddenly, are brief, and
are often disturbing may emphasize the urgency of solving a potential prob-
lem. It is also conceivable that intrusions themselves present as a mechanism
to process information, in line with Horowitz. Here, we begin by examining,
in nonclinical populations, possible feelings that may be associated with intru-
sive thoughts, and we address their potential adaptive mechanism (intrusion
content). Thereafter, we approach the possibility that intrusive thoughts in and
of themselves are a mechanism that enables working through unaddressed but
identifiable problems by viewing intrusions through the lens of memory (intru-
sion process).
Intrusion Content: Conducive Nonclinical Instances
Though they are not necessarily centered around a disorder, intrusions in non-
clinical samples appear to occur along common themes across individuals. We
contend that these themes may provide insight into the potential adaptive na-
ture of intrusions, in line with the adaptationist framework proposed above.
The following list is not exhaustive, but provides a starting point to examine
different forms of emotions that can be at the source of intrusions. Their com-
mon characteristic is that they might serve an adaptive purpose; however, de-
pending on the context during which they occur, intrusions can also reflect an
adaptation gone awry.
Anger, Rage, and Revenge. Anger has been hypothesized to be an evolved

emotion, the expression of which functions (in part) to recalibrate someone
else’s welfare trade-off ratio with respect to you (Sell 2011). When a behav-
ior affects two or more individuals, it can be selfishly skewed or altruisti-
cally skewed. Consider a roommate who has left dirty dishes strewn about
the shared kitchen, expecting you to clean them. Expressions of anger to
the roommate communicate that they have insufficiently taken your welfare
into account and should adjust it in the future. In this simple example, in-
trusive thoughts and prolonged rumination function to stoke the emotion of
anger until the roommate arrives back home and the rage can be expressed,
ideally causing the roommate to recalibrate their welfare trade-off ratio with
respect to you.
Now consider road rage: When someone cuts you off in traffic, it some-
times activates intense anger. In the modern environment, road rage sometimes
produces violent car accidents when ramming the violator. The underlying
emotion evolved presumably in small-group contexts in which its expression
would cause the violator to recalibrate, taking your welfare more into consid-
eration in the future. In a modern environment marked by dense urban living
patterns, in which the handling of severe social violations has been outsourced
to professional police, expressions of road rage can lead to disastrous and mal-
adaptive outcomes (e.g., car crashes, personal injury, and death). The design
feature of intrusive thinking that prolongs rumination about the violator was
presumably adaptive in small-group contexts of the past, where social reputa-
tions mattered greatly and the failure to respond to violations could lead to a
catastrophic loss of status. In the modern mismatched environment contain-
ing lethal 4,000-pound vehicles, traffic congestion, and swarms of anonymous
strangers, intrusive rumination about someone who cut you off can lead to road
rage and a maladaptive misfiring of this ancient emotion.
Jealousy and Infidelity. Jealousy is an emotion that evolved to combat threats

to a valued social relationship. If the relationship is a mateship, jealousy can
be activated by cues to sexual or emotional infidelity, to signs of a partner’s
defection, or to threats posed by potential mate poachers or even by mate value
discrepancies (Buss 2000; Buss and Haselton 2005). Infidelity is typically
cloaked in secrecy, creating a signal detection problem for the partner. Once
jealousy is activated, it can produce intrusive thoughts, prolonged rumination,
and motivate vigilance to discern the nature and magnitude of the relationship
threat. Intrusive thinking in this context can be functional, leading a person
to gather relevant information and to allocate effort to warding off the threat,
devoting resources to mate retention or to repelling the genuine threat posed

by a potential mate poacher.
Intrusive thinking is a design feature of the jealousy adaptation; it leads its
bearers to uncover and attempt to solve real threats to romantic relationships.
Nonetheless, it can also misfire, leading to maladaptive outcomes. If the psy-
chological detection of infidelity cues is set too sensitively, it can produce false
accusations of infidelity, undermining the very relationship that jealousy was
designed to protect. It can produce delusions of a partner’s infidelity and patho-
logical jealousy, leading to extreme violence toward a partner (Buss 2000).
Because infidelities are typically concealed, cues to infidelity are inherently
probabilistic. Based on error management theory, there is evidence that people
overinfer infidelity to avoid the costly error of losing a partner to a roman-
tic rival, even at the cost of making more frequent errors of inference (Goetz
and Causey 2009). Moreover, many individuals who have been diagnosed by
psychiatrists as having delusional or pathological jealousy turn out to have
partners who, upon deeper investigation, have actually been unfaithful (Buss
2000). In short, it is difficult in any particular case to determine unambiguously
whether jealous intrusive thinking is functioning as it was designed to func-
tion, or if it is misfiring and causing pathological outcomes.
Love and Romantic Infatuation. Intrusive thinking is a common feature of

the infatuation stage of love, markedly present when separated from a loved
one (Fisher 2016). It can interfere with work, cause other relationships to lapse,
and even create a metabolic deficit when someone forgets to eat. Intrusive
thinking often creates an idealization of the loved one, imputing maximal val-
ues to desirable qualities that have not yet been observed. Preoccupation pre-
sumably leads to efforts to woo a loved one or to become reunited with them
after separation. After the infatuation stage fades and is replaced by a more
subdued warmth and attachment, intrusive thinking subsides, allowing a real-
location of effort to other adaptive challenges, such as obtaining food, negotiat-
ing status hierarchies, or solidifying coalitional alliances. Intrusive thinking in
the context of the infatuation stage of love is temporally delimited.
Like all adaptations, this one can go awry, misfiring in the modern envi-
ronment. People develop romantic infatuations with movie stars, for example,
when there is no possibility of meeting them, much less successful consum-
mation. In the extreme, these can lead to criminal stalking, as in the case of
John Hinckley Jr. who developed an intense infatuation with the actress Jodie
Foster. He sent her numerous love letters, stalked her, and when his efforts
failed to produce reciprocation, he attempted to assassinate President Ronald
Reagan in a last-ditch desperate attempt to get her attention and demonstrate
the intensity of his love and commitment to her. He now resides in a prison
cell. In short, intrusive thinking can lead to disastrous outcomes, both for the
individual and for others who become victims. When properly functioning,
however, intrusive thinking leads to successful consummation of love.
This illustrates that intrusive thinking is not solely a design feature of the
so-called “negative emotions.” It is likely an evolved design feature of many
emotions, including fear, rage, jealousy, shame, and guilt, as well as more
positively valenced emotions such as love and sexual arousal. Nor is it al-
ways dysfunctional. Intrusive thinking is often a key design feature, motivat-
ing attention to pressing adaptive problems while postponing effort allocated
to less important ones. Error management analysis highlights the difficulty of
distinguishing functional from dysfunctional outcomes in any specific case,
rendering the theoretical analysis of intrusive thinking more complicated than
previously considered. Another example that generally involves positive in-
trusions pertains to the pursuit of goals or aspirations. It is important to note
that while some emotional contexts are conducive to either positive or nega-
tive intrusions, others are likely to be more complex, as in the case of grief.
Intrusions can often be autobiographical, that is, they relate to an individual’s
firsthand experience. Cast in this light, intrusions can actually be interpreted as
a memory retrieval in some instances. We briefly consider intrusions occurring
in the context of goals or aspirations and grief, and then further examine intru-
sions as memory retrieval to extract their potential adaptive process.
Goals or Aspirations. Intrusions can occur in scenario building of means to

achieve. In such a case, the intrusions would likely be positive and inspire some-
one to pursue achievements. Much like the cases described above, the adaptive
nature of goal and aspiration intrusions can go awry. For instance, an individual
may experience fantasy-like intrusions that reach far beyond their abilities. In
this case, the originally positive intrusions could grow to be a reminder of one’s
failures and hinder one’s potential to succeed in a more achievable realm.
Grief. Grief is almost always triggered by the loss of a key social partner—a
close friend, a romantic partner, or a family member. Research on intrusions dur-
ing grief is limited but suggests the presence of both positive and negative intru-
sions in individuals experiencing the loss of a loved one (Boelen and Huntjens
2008). In the positive realm, mourners may experience intrusive memories of
the loved one that died or fantasy reenactment. Through the lens of memory
(discussed below), the purpose of grief intrusions could be that of strengthening
a neurobiological trace, to keep the memory alive. Negative intrusions of grief
can include memories of the death event or negative images or thoughts about the
future. Early during grief, the intrusions, both positive and negative, may be help-
ful to the individual who experienced a loss. However, if persistent and enduring,
they could interfere with a person’s ability to move forward.
Evolutionary scholars have advanced two competing explanations of grief.
One is that grief is an unfortunate nonadaptive by-product of love and attach-
ment (Archer 2003), both of which are profoundly important adaptations in
the evolved social suite of humans (Christakis 2019). The second is that grief
serves several adaptive functions, such as identifying actions that might have
led to the loss, motivating actions to prevent future losses and signaling to sig-
nificant others (friends, family, mates) the need for help due to the loss (Nesse
2005). Other possible functions include signaling to others that you are a loyal
coalitional ally and ruminating about the implications of the loss for replacing
the lost one with an alternative mate or coalitional ally. Which of these compet-
ing hypotheses, or which combination, will bear fruit rests with future empiri-
cal research on individuals who experience loss and grief.
Intrusion Process: Intrusions as Memory Retrieval
There are useful parallels to be drawn between intrusions and other forms of
memory retrieval. Memory retrieval can be broadly defined as recalling a prior
experience, either following the presentation of an external or internal cue or
through volitional control. Conceptualizing intrusions as memory retrieval ap-
pears very much in line with Horowitz’s definition of intrusive thoughts and
enables approaching the concept with an adaptive mechanistic view.
From this perspective, we can think of intrusions as potentially serving the
adaptive functions described below. We can also conceive of intrusions as pro-
viding an opportunistic window or intervention. The latter can best be under-
stood through the process of reconsolidation and memory updating.
Reminder That Certain Information Needs to Be Further Processed. By vir-

tue of being interruptive and often irrational, intrusions are noticed. In this
case, intrusions would likely reflect an event that has passed, which an indi-
vidual may need to prioritize or address.
Warning to Allow Preparedness. Intrusions draw attention. Their purpose

here is to enable an individual to react in the presence of a looming situation.
As such, the intrusions could include content related to an individual’s past
experiences, but would certainly pertain to an individual’s future.
Mechanism to Initiate Extinguishing or Exerting Another Form of Inhibitory

Control over a Negative Memory. Retrieval of a previously consolidated
memory (i.e., a memory that has been stored into long-term storage for longer
than ~six hours) engages two seemingly opposing mechanisms: reconsolida-
tion and extinction. Reconsolidation refers to a putative process which proposes
that after retrieval, previously consolidated memories become destabilized and
require renewed protein synthesis for long-term storage. Reconsolidation also
offers an opportunistic window during which memories can be updated. In
extinction, the repeated presentation of the conditioned stimulus in the absence
of the unconditioned stimulus leads to a progressive decrease in the behavioral
expression to the stimulus. Extinction can refer to both a process (the progres-
sive decrease in fear throughout a session) and an outcome (e.g., resultant de-
crease in fear responding).
The concepts of reconsolidation and extinction have been extensively stud-

ied, and each provides an important avenue of improving psychotherapeutic
outcome, particularly in anxiety-related disorders and addiction (Monfils and
Holmes 2018). The two approaches have also been successfully combined to
improve upon long-term therapeutic outcomes in the retrieval-extinction para-
digm (Monfils et al. 2009). Approaching intrusions as memory retrieval could
potentially enable the optimization of therapeutic approaches. If an intrusion
is mechanistically akin to memory retrieval, it could provide an opportunistic
window to intervene and attenuate their potency. Effectively, research suggests
that behavioral or pharmacological interventions shortly after memory retrieval
improves outcome above and beyond standard extinction-based approaches
(Monfils et al. 2009; Schiller et al. 2010; James et al. 2016b; Telch et al. 2017).
Another way to handle upsetting memories, once recalled, is to exert a form
of inhibitory control over them (other than extinction). A number of such ap-
proaches are discussed in detail by Visser et al. (this volume).
Memory Strengthening or Maintenance. Once they are retrieved, memories

generally strengthen if left untargeted (Inda et al. 2011). As such, while an
intrusion may present a window of opportunity for treatment of a traumatic
memory, if untreated, the intrusion could actually lead to memory strength-
ening as well. Such a mechanism could provide an adaptive explanation for
certain intrusions (e.g., those that manifest during grief, or during goals and as-
pirations). In other cases (e.g., following trauma), intrusions could exacerbate
a negative memory and render treatment more challenging.
Means of Escaping Boredom. While there are often specific circumstances

or conditions that appear to prime the presence of an intrusion, others may be
more random. In such a case, an intrusion could conceivably serve the adap-
tive purpose of escaping boredom or monotony in a safe way by engaging in a
daydream experience. In this context, intrusions could, for example, promote
an individual to engage in mind wandering. An extensive discussion of mind
wandering can be found in Visser et al. (this volume).
Summary
Intrusions are what we make of them. Although sometimes acutely distressing

when they occur, intrusions can, in and of themselves, actually be innocuous
or even positively valenced. What appears to be at the source of most of the
distress experienced by intrusions is often the thought process that follows.
Consider the following scenario:
A person is standing on the subway platform listening to their favorite song
with headphones. Suddenly, a vivid intrusion appears in their mind: they see
themselves jumping on the tracks, just as a train passes through. Thus far, the
intrusion has not per se caused any harm. The person’s reaction to the intru-
sion, however, can vary. One individual could simply think: “Whoa! That was
crazy! Of course, I would never do that.” At the same time, this person might
experience a sense of comical relief: “Boy, if I had jumped, I wouldn’t have
to sit through all of those blankety-blank-blank meetings scheduled today.”
Another person might process the intrusion differently. For instance, they
might think: “Why am I having these images? Should I jump? I am worthless.
I don’t want to die. Or maybe I do want to die. I don’t know what to do. I’m
worried about what I might do.”
In the first response scenario, the person may not be bothered further (or
at all) by the intrusion. In the second case, a person might perseverate on the
experience or engage in behaviors to try to minimize the impact of the intru-
sion. This could potentially result in worry, rumination, and/or obsessions and
associated compulsions.
In a sense, intrusions themselves may not be as distressing as what we make
of them, and what we make of them is likely to be largely influenced by our
state of mind (including, in clinical manifestations, the underlying pathology).
In approaching treatment for individuals who experience intrusions, it is im-
portant to consider their possible adaptive nature. In doing so, it may be helpful
to identify the intrusions’ possible underlying content as well as the psycho-
logical process that a person’s brain has determined should be engaged via the
intrusive thought. Ultimately, identifying the possible “why” of intrusions may
help guide “how” we can best treat them (Figure 7.1).
8
Current Psychiatric Perspectives

on Intrusive Thinking
Florian Schlagenhauf, Andreas Heinz, and Martin Voss
Abstract
Various psychopathological symptoms share characteristics of intrusive thinking. Intru-
sive thoughts are part of the diagnostic criteria for posttraumatic stress disorder and ob-
sessive-compulsive disorder but are also relevant in other psychiatric conditions, such
as drug craving in addiction or rumination in depressive disorders. Intrusive thoughts
must be differentiated from thought insertion observed in schizophrenia and related
psychotic disorders. This chapter reviews the typical characteristics and content of in-
trusive thinking in the context of different psychiatric conditions and outlines current
theories regarding the mechanisms of intrusive thinking.
Introduction
Intrusive thoughts can be characterized as repetitive, uncontrollable, distress-

ing thoughts that enter conscious awareness unwantedly (Clark 2005). They
are an important aspect of different psychiatric disorders, but they also mani-
fest in the nonclinical, general population (Clark 2005; Garcia-Soriano et al.
2011). Prominent psychiatric examples include intrusions related to traumatic
events in patients suffering from posttraumatic stress disorder (PTSD) as
well as aggressive obsessions experienced by obsessive-compulsive disorder
(OCD) patients (Heinz 2017). In schizophrenia or related psychotic disorders,
the delusions, hallucinations, or thought insertions experienced by patients are
repetitive, uncontrollable, and distressing (Heinz et al. 2016). In each of these
conditions, the content of intrusive thoughts can be very different: in PTSD,
the content may refer to a real autobiographic event whereas in OCD it may
relate to an obsessive thought about contamination. Certain similarities do,
however, exist: characteristic thoughts appear repeatedly (to a degree) and can
interfere with normal functioning. A person perceives the thoughts to be un-
wanted and reports having no control over these thoughts; that is, they do not
result from a deliberate or even effortful process, but appear involuntarily and
114 F. Schlagenhauf, A. Heinz, and M. Voss
automatically. Finally, intrusive thoughts create distress due to their content

and/or characteristics.
In this chapter, we discuss intrusive thinking in the context of PTSD, OCD,
addiction, and schizophrenia. For each of these psychiatric conditions, we out-
line the typical characteristics and content of intrusive thinking, together with
the diagnostic criteria of the respective psychiatric condition. Discussion fol-
lows on current theories that aim to explain the mechanisms behind intrusive
thinking, and we conclude by reviewing problems and open questions that re-
quire future attention.
Intrusive Thoughts in Psychiatric Conditions
Posttraumatic Stress Disorder
One prominent and required diagnostic criterion for PTSD is the presence of
intrusion symptoms: a traumatic event is persistently reexperienced. According
to DSM-5, Criterion B (American Psychiatric Association 2013), these intru-
sion symptoms encompass the following characteristics:
• Recurrent, involuntary, and intrusive memories of the traumatic event(s)
• Traumatic nightmares
• Dissociative reactions (e.g., flashbacks)
• Intense distress after being exposed to traumatic reminders
• Heightened physiologic reaction to trauma-related stimuli
To meet the full diagnostic criteria, a person must have been exposed to a life-
threatening event and thereafter has to avoid trauma-related stimuli. Further,
the person exhibits negative alterations in cognition and mood, which began
or worsened after the traumatic event, and displays alterations in arousal and
reactivity, such as hyperarousal.
The clinical characteristic of intrusive memory is that it “springs to mind
unbidden—that is, against the person’s will” (Visser et al. 2018). Such intru-
sive memories are forms of episodic memories of actually experienced auto-
biographical events, which are retrieved involuntarily. In its extreme form, the
person intensely and vividly relives the traumatic event in the present. Such
flashbacks involve the retrieval of detailed sensory features and are highly
emotional. Typically, fragments and several distinct moments of the trauma
are recalled, the so-called “hot spots,” in a predominantly visual form (Visser
et al. 2018).
Disturbance in memory seems to be the prominent feature of PTSD. It is
widely agreed that multiple memory systems exist and that these rely, in part,
on distinct neurobiological substrates (Henke 2010). “Declarative” memories
are events and facts, which we can explicitly remember, and seem to depend on
medial temporal lobe structures. “Nondeclarative” memories are implicit and
Current Psychiatric Perspectives on Intrusive Thinking 115
not consciously accessible. They encompass procedural memory (e.g., aver-

sive conditioning, motor skills, and habits) and are thought to be subserved by
subcortical areas such as the amygdala (aversive conditioning) and striatum
(skills and habit formation).
To what degree do different aspects of intrusive aversive memories (e.g.,
visual imagery or physiological responses) relate to different memory sys-
tems? Physiological reactions to trauma-related cues, also listed as intru-
sive symptoms in DSM-5, Criterion B (American Psychiatric Association
2013), are nondeclarative memories triggered by stimuli. Conversely, “un-
wanted emotion-laden memories that spring to mind unbidden in the form
of sensory imagery” (Visser et al. 2018; see also Figure 14.1, Holmes et
al., this volume) belong to the declarative system. Therefore, different neu-
robehavioral mechanisms most likely underpin the heterogeneous intrusive
symptoms that are observed in PTSD. It has been suggested that interven-
tion strategies should specifically target involuntary rather than voluntary
retrieval (Visser et al. 2018), both intrusive memory fragments of a trauma
as well as the conditioned responses to trauma-related cues are experienced
involuntarily. Thus, a successful therapeutic intervention should aim to
modify the specific underlying memory traces, while preserving an indi-
vidual’s ability to deliberately recall episodes and facts about the trauma
(e.g., for legal reasons).
Obsessive-Compulsive Disorder
The defining features of OCD are repetitive, distressing, and inappropriate

thoughts (obsessions) and/or actions (compulsions). According to DSM-
5, obsessions are “recurrent and persistent thoughts, urges, or images that
are experienced…as intrusive and unwanted, and that in most individuals
cause marked anxiety or distress” (American Psychiatric Association 2013).
Common foci of obsessions include contamination, pathological doubt,
need for symmetry, and aggressive or sexual content. Patients fail to ignore
or suppress obsessions and instead attempt to neutralize them through other
thoughts or actions, such as by compulsively performing ritualistic behavior
to undo the alleged harm of the obsessive thought or intention. Compulsions
are thus repetitive behavioral or mental acts, such as checking, washing, or
counting. These acts aim to reduce anxiety or distress, but they lack a realistic
connection between the act and the goal that a person should achieve. Patients
are aware that their compulsions and obsessions are unreasonable and inap-
propriate. The symptoms are experienced as alien and disturbing (i.e., ego-
dystonic); still, they are recognized as being caused by the afflicted patient and
not by an external agent (Heinz 1999).
Cognitive theories state that dysfunctional beliefs are the core of OCD and
that compulsions develop to reduce anxiety. Appraisal models, for instance,
posit that subjects who appraise intrusions as significant and meaningful (based
on their dysfunctional beliefs), can develop OCD by escalating intrusions into

obsessions (Julien et al. 2007). This hypothesis is based on the observation
that nonclinical individuals experience intrusive thoughts, images, or impulses
similar in content to individuals with OCD. Other recent theories conceptual-
ize OCD as a disorder of habitual control (Robbins et al. 2019), where post
hoc rationalizations of habitual actions contribute to obsessions (Gillan and
Robbins 2014). Dual-system theories state that human learning and adaptive
behavior are governed by two interacting control systems: one mediates goal-
directed actions while the other supports habits (Balleine and O’Doherty 2010;
Dolan and Dayan 2013). Habits are performed autonomously and largely in-
dependent of their consequences and are, therefore, inflexible to changes in
reward contingency. In contrast, goal-directed actions are performed because
of anticipated outcomes; thus, they rely on forward planning and allow greater
flexibility to changes in contingencies (Friedel et al. 2014). There is a strik-
ing similarity between compulsions and habits (Gillan and Robbins 2014):
Compulsions are automatic behaviors experienced as irrational and not in line
with current goals (i.e., they are ego-dystonic). Behaviors characterized as hab-
its are insensitive to action–outcome contingency and outcome value. Intrusive
thoughts are perceived as unintended and not as deliberate and instrumental
mental acts; this suggests that a more implicit and nondeliberative way of in-
formation processing is involved in intrusive thinking, more akin to a habitual
system than a goal-directed control system. Thoughts can be accompanied by
sensory qualities (e.g., intrusions in PTSD), or they can be heard aloud, which
traditionally has been distinguished from thought insertion and classified as
acoustic hallucinations (Jaspers 1946).
The temporal sequence of OCD symptoms seems relevant for an etiologi-
cal understanding of OCD: Do obsessions come first and compulsive acts fol-
low to reduce negative emotional states, as put forth by the cognitive theories
of OCD? Alternatively, do compulsions develop first such that obsessions are
secondary rationalizations of the compulsions? There seems to be limited lon-
gitudinal data to answer these questions reliably. However, a large proportion
of children with OCD deny that their compulsions are driven by obsessive
thoughts (Robbins et al. 2019). On the other hand, adult OCD patients report
that intrusive images of dirtiness or contamination evoke the urge to wash or
neutralize (Coughtrey et al. 2012).
Intrusive thoughts and compulsions present in OCD have been linked to
dysfunction of frontostriatal circuits (orbitofrontal/anterior cingulated cortex,
dorsolateral striatum/caudate, thalamus) (Robbins et al. 2019). Several neu-
roimaging studies of OCD revealed hyperactivity in these brain areas during
rest (e.g., Baxter et al. 1987) and cognitive performance (e.g., van den Heuvel
et al. 2005). Similar neurocircuits were activated during symptom provoca-
tion, mainly using visual stimuli to trigger OCD symptoms (Breiter and Rauch
1996). Moreover, tonic overactivity of this circuit seems to be associated
with symptom severity (Adler et al. 2000) and predicts treatment response.
Frontostriatal dysfunction normalizes after successful treatment with psycho-

tropic medication (e.g., Swedo et al. 1992) or cognitive behavioral therapy
(e.g., Nakao et al. 2005).
In light of these neurobiological findings, Heinz (1999) suggested that
dopamine dysfunction in the dorsal striatum is associated with motor tics,
whereas more complex compulsive behavior patterns are triggered by cogni-
tive concerns processed in the orbitofrontal cortex, which persist obsessively
due to impaired feedback processed in dorsal striatal-thalamic-frontocortical
loops. This neurocircuit model may help unify current theories that focus on
cognitive versus habitual aspects of OCD.
Depressive Disorders and Rumination
Rumination is a maladaptive form of self-reflection and recursive self-

focused thinking. Like obsessions, rumination involves recursive thinking
about particularly self-centered negative information. However, whereas
obsessive thoughts are usually experienced as aggressive or otherwise in-
appropriate (and are hence “unwanted” by the afflicted person), rumination
often focuses on threatening environmental conditions as well as inappropri-
ate or unfavorable character traits of the person (and thus involve self-blame
for not being able to cope with the situation). Rumination has been hypoth-
esized as an important factor in developing depressive symptoms and has
been shown to exacerbate depression, enhance negative thinking, and impair
problem solving (Nolen-Hoeksema et al. 2008). Self-focused rumination in
depressive disorders has been linked to self-referential processes and the
brain’s default-mode network, particularly involving the subgenual prefron-
tal cortex, a brain area implicated in the processing of aversive information
and the modulation of negative mood states (Kühn et al. 2013; Hamilton et
al. 2015). Accordingly, positron emission tomography studies have revealed
increased metabolic activity in this brain area among subjects with major
depression (Drevets et al. 2008).
Major depression and OCD appear to differ with respect to their neuro-
biological correlates, which may reflect differences in relevant pathological
mechanisms: In major depression, rumination may result from a failure to reg-
ulate aversive information input and associated personal concerns. In OCD,
compulsive behavior appears to be triggered by obsessive thoughts aimed at
compensating for the aggressive or otherwise unwanted content of these ob-
sessions, yet fails to dampen concerns, which triggers repetitive action, re-
sulting in reverberating circular interactions. In depressive disorders, other
instances of intrusive thinking include suicidal ideations; mental images of
killing oneself (“flash-forward” thoughts) have been shown to be associated
with suicidal behavior (for further examples of mental imagery, see Holmes
and Mathews 2010).
Addiction: Craving and Compulsivity
Intrusive thinking in addiction disorders includes thoughts related to drug con-

sumption associated with a strong desire (craving) to consume a drug despite
having reached a conscious decision to abstain from drug use (Heinz 2017).
In this context, craving is defined as a strong desire or urge to use a drug or to
engage in harmful activity, such as gambling for monetary reward. Further cri-
teria for addiction include impaired control regarding substance intake, neglect
of unrelated activities, tolerance, and withdrawal. Despite conscious decisions
to do otherwise, recurrent substance use is a key characteristic of addiction,
and it has been suggested that drug use becomes compulsive when subjects
lose control over the powerful urge to consume a drug of abuse despite of
aversive consequences (Everitt and Robbins 2016).
Indeed, a gradual shift from outcome-sensitive, goal-directed behavior to
habitual behavior can contribute to automatic, habitual, or even compulsive
drug intake, despite foregone positive outcomes and devastating (future) nega-
tive consequences. Drug-associated stimuli may acquire enhanced salience
and act as appetitive Pavlovian cues that trigger automatic approach behavior
(Robinson and Berridge 1993). Moreover, these environmental cues can impact
(goal-directed) choice selection and behavioral adaptation through Pavlovian-
instrumental transfer mechanisms, where affectively positive Pavlovian cues
bias (unrelated) goal-directed behavior toward approach even when this is not
useful in the instrumental context (Garbusow et al. 2016).
While healthy controls are able to arbitrate control between the habitual
and the goal-directed system, a loss of control over certain behaviors (e.g.,
drug intake) might be due to a shift from goal-directed toward habitual control
(Balleine and O’Doherty 2010; Dolan and Dayan 2013; Voon et al. 2015).
Computational neuroscience uses “model-based” and “model-free” algorithms
to explain goal-directed and habitual learning during, for example, sequen-
tial decision-making tasks (e.g., Friedel et al. 2014). A model-based algorithm
views the environmental (or task) structure used for deliberative forward plan-
ning as a hallmark of goal-directed behavior, which in the case of sequential
decision making refers to the transition from one environmental state to an-
other. Model-free reinforcement learning algorithms reflect a retrospective and
more rigid strategy that neglects environmental structures and relies solely on
repeating previously rewarded actions. Initial studies in patients with different
addictive disorders, including dependence on psychostimulants and alcohol,
suggest impaired model-based control, thus shifting the behavior toward a
model-free response (Sebold et al. 2014; Voon et al. 2015). However, in a more
recent study, Sebold et al. (2017) found neither a general bias toward model-
free (supposedly habitual) decision making in patients with alcohol dependence
nor a poor treatment outcome associated with impaired model-based deci-
sion making. Instead, the balance between habitual (supposedly model-free)
and goal-directed decision making differentiated alcohol-dependent patients
(who later relapsed) from abstainers and controls only when individual alco-
hol expectancies were considered. This suggests that habitual behavior does
not generally increase in addicted patients during choice behavior tests for
non-drug-related rewards in a laboratory context; addiction-related habits ap-
pear to be triggered by specific cues and contexts in conjunction with previ-
ous experiences. In line with this hypothesis, a patient suffering from OCD,
pathological gambling, and drug addiction described constant urges to perform
habits compulsively related to his obsessions, while craving for gambling and
drug intake was triggered only during certain time periods by specific drug or
gambling-related stimuli (Schoofs and Heinz 2013). In light of these findings,
compulsions in OCD appear to differ significantly from “compulsive” urges to
consume drugs of abuse, warranting further phenomenological and neurobio-
logical specifications (Heinz 2017).
Schizophrenia and Related Psychotic Disorders and Thought Insertions
Thought insertion is a positive symptom of schizophrenia and is regarded as

a “first rank symptom” of the disease (Schneider 1959; Heinz et al. 2016).
Not only does thought insertion constitute one of the most astounding posi-
tive symptoms of schizophrenia, it is frequently expressed. It occurs in ap-
proximately half of all patients diagnosed with schizophrenia (Sartorius et al.
1977), but appears to be absent in organic psychoses (Marneros 1988; Heinz
et al. 1995).
In psychosis, patients typically report that thoughts are being “inserted” (in
verbal form) by another agent into their head. Patients thus lose the feeling of
“mineness” for a given thought; this marks a distinct difference between obses-
sions in OCD, ruminations in major depression, or drug cravings in addiction,
all of which are “unwanted” and uncontrolled by the afflicted subject but not
experienced as “alien” and attributed to outside agents. Vosgerau and Voss
(2014) highlight the distinction between control, ownership, and authorship of
thoughts. They argue that it is a conceptual truth that introspected thoughts are
necessarily owned by the introspector (therefore ownership of thoughts cannot
be disturbed), whereas lack of authorship over thoughts can be experienced
in everyday phenomena (thinking “communicated thoughts,” i.e., thoughts
clearly formulated by another person) as well as in pathological conditions
such as psychosis. By introducing another factor (e.g., control over thoughts),
Vosgerau and Voss (2014) argue that the phenomenon of thought insertion is
caused by a combination of these two factors—lack of control and lack of au-
thorship—and that there is a double dissociation between both factors.
In an attempt to reveal the neurocognitive mechanisms underlying thought
insertion, Campbell (1999) drew an analogy between thoughts and motor con-
trol processes and explained thought insertion in relation to the comparator
model, originally developed for motor control (Frith et al. 2000). Campbell
assumed that thoughts are comparable to motor processes (similar views were
expressed by Feinberg 1978 and Ito 2008) and that every thought is preceded
by an intention to think this thought. The actual thought occurring in one’s
stream of consciousness is then compared with the intention to think. When
these two processes match, a feeling of authorship results; when they do not,
attribution to another agent may occur.
While this offers an appealing framework to explain passivity phenomena,
such as delusions of control or hallucinations, several problems arise when the
comparator model is used to explain thought insertion (for a detailed critique,
see Vosgerau and Synofzik 2010). One problem is that the account does not
distinguish thinking (as a process to generate thoughts) from thoughts (as a
result of thinking), making it difficult to pinpoint the difference between “in-
fluenced” thinking and “inserted” thoughts. Furthermore, it remains unclear
what an intention to think a specific thought could be and how it can be dis-
tinguished from the actual thoughts; that is, why the intention to think is not
naturally conceived of as the thought itself. Indeed, if every thought were to
be preceded by an intention to think, we would run into an infinite regress: for
each thought, we need a thought to get the process started, which in turn pre-
supposes another thought, and so on.
A more recent attempt to conceptualize the phenomenon of thought inser-
tion treats inserted thoughts as sensory events rather than motor processes
(Sterzer et al. 2016). Building on detailed phenomenological descriptions
from the early Heidelberg school in the first half of the twentieth century,
which described thoughts that “become sensory” as being experienced as in-
serted, more recent studies explain thought insertion within the framework of
predictive coding and Bayesian inference (Sterzer et al. 2016). Here, thought
insertion is viewed not as the failure of introspection in a comparator process
(Campbell 1999) but rather as the failure (or imprecision) of prior beliefs
in a Bayesian inference process thought to be at the core of thought inser-
tion (Sterzer et al. 2016). In analogy with aberrant salience attribution to
external events, which could lead to the emergence of delusional mood and
fixed beliefs (Heinz 2002; Kapur 2003; Heinz and Schlagenhauf 2010), in-
ternal events (e.g., verbalized thoughts) may also be experienced as overly
salient, and therefore unusual, as well as surprising due to a lack of context
and unusual structure (with a possible link to formal thought disorder). The
individual’s attempt to explain the aberrant salience and unusual character of
such verbalized thoughts could result in their interpretation as being exter-
nally caused. Nonetheless, unintended or semantically inappropriate verbal-
izations can also be expressed in aphasia (e.g., due to a stroke) but are not
accompanied by reports of “alien” involvement (Heinz 2017). Therefore, ad-
ditional steps may be required to convince a person that a thought is “alien”
and thus “must” be inserted by an external agent. In this context, it has been
suggested that low precision of prior beliefs and/or increased sensory preci-
sion, both inaccessible to introspection, may render some thoughts so unpre-
dictable that they are experienced as inserted (Heinz et al. 2019). Whereas
beliefs and desires may have a role in prior beliefs regarding our own
thoughts (Stephens and Graham 2000), what makes a thought feel alien is
not that they are simply unwanted or “immoral” (in the sense of higher-level
introspection), but rather that patients directly notice or perceive a thought
to be “alien,” thus pointing to unconscious mechanisms causing this experi-
ence. Such mechanisms may be usefully described by a Bayesian account of
information processing in the central nervous system (Adams et al. 2013):
during psychotic episodes, if prior beliefs are indeed imprecise in compari-
son to sensory input-driven posteriors, frequent prediction errors are made,
which may trigger phasic dopamine release and cause salience to be attrib-
uted to otherwise irrelevant cues, including verbalized thoughts (Sterzer et
al. 2016; Heinz et al. 2019).
If these considerations are correct, they may help to explain the difference
between “alien” thought insertion and other forms of unwanted or intrusive
thought content. In psychosis, impaired precision of prior beliefs may affect
the whole experience of the world and self, which are experienced as unusual,
alien, and often threatening. Thus, thought insertion in psychosis goes beyond
the experiencing of unusual verbalizations (as in aphasia), unwanted cravings
(in addiction), self-centered concerns (in major depression), intrusive memo-
ries (in PTSD), or obsessive thoughts (in OCD). The entire relationship be-
tween the individual and the real world is affected: that which was well known
for a long time suddenly carries hidden meanings, harmless situations are im-
bued with a sense of danger, and longtime friends and family members become
deeply alienated and may no longer be trusted. Since Bayesian accounts are
supposed to reflect general functions of the central nervous system, such com-
putational frameworks will have to explain how more restricted alterations in
information processing, particularly with respect to verbalized thoughts, differ
from the more fundamental alterations experienced in psychotic states.
Open Issues
To guide future enquiry, we conclude our discussion by highlighting unsolved

problems that await clarification through future research. First, despite strik-
ing phenomenological differences between negative verbal thoughts, intru-
sive visual images, and memories, we need to know whether the underlying
psychological and neurobiological mechanisms involved in intrusive thinking
are similar across diagnostic categories. In support of shared transdiagnostic
mechanisms, Gillan et al. (2016) has shown that reduced goal-directed control
is associated with compulsive behaviors and intrusive thoughts.
Second, the relation between intrusive thinking and the concept of compulsiv-
ity needs to be elucidated, where compulsivity is defined as “a hypothetical trait in
which actions are persistently repeated despite adverse consequences” (Robbins
et al. 2012:82). Are intrusive thoughts one manifestation of compulsivity? If so,
when is it expressed in maladaptive actions and when in intrusive thoughts? Can

intrusions be understood as mental habits?
Finally, in psychosis, how can we mechanistically isolate disturbed au-
thorship (or disturbed “mineness”) from disturbed control over thoughts?
Importantly, measurement instruments have to be harmonized on the clinical,
psychopathological, behavioral, and neurobiological levels.
9
Neuropsychological
Mechanisms of
Intrusive Thinking
Renée M. Visser, Michael C. Anderson, Adam Aron,
Marie T. Banich, Kathleen T. Brady, Quentin J. M. Huys,
Marie-Hélène Monfils, Daniela Schiller, Florian Schlagenhauf,
Jonathan W. Schooler, and Trevor W. Robbins
Abstract
A classic definition of intrusive thinking is “any distinct, identifiable cognitive event
that is unwanted, unintended, and recurrent. It interrupts the flow of thought, inter-
feres in task performance, is associated with negative affect, and is difficult to control”
(Clark 2005:4). While easy to understand and applicable to many cases, this definition
does not seem to encompass the entire spectrum of intrusions. For example, intrusive
thoughts may not always be experienced as unpleasant or unwanted, and may in some
situations even be adaptive. This chapter revisits the definition of intrusive thinking, by
systematically considering all the circumstances in which intrusions might occur, their
manifestations across health and disorders, and develops an alternative, more inclusive
definition of intrusions as being “interruptive, salient, experienced mental events.” It
proposes that clinical intrusive thinking differs from its nonclinical form with regard to
frequency, intensity, and maladaptive reappraisal. Further, it discusses the neurocogni-
tive processes underlying intrusive thinking and its control, including memory pro-
cesses involved in action control, working memory and long-term memory encoding,
retrieval, and suppression. As part of this, current methodologies used to study intrusive
thinking are evaluated and areas are highlighted where more research and/or technical
innovation is needed. It concludes with a discussion of the theoretical, therapeutic, and
sociocultural implications of intrusive thinking and its control.
Group photos (top left to bottom right) Renée Visser, Michael Anderson, Marie-
Hélène Monfils, Trevor Robbins, Daniela Schiller, Marie Banich, Adam Aron,
Kathleen Brady, Jonathan Schooler, Florian Schlagenhauf, Quentin Huys, Michael
Anderson, Daniela Schiller, Jonathan Schooler, Kathleen Brady, Quentin Huys, Florian
Schlagenhauf, Marie-Hélène Monfils, Trevor Robbins, Renée Visser, Marie Banich
126 R. M. Visser et al.
Introduction
While reading this chapter, you might be sitting in a coffee shop somewhere
or traveling on a train. The sunlight comes in through the window and you
envision how warm the planet might get. You are having a meeting in a few
hours and various scenarios of it are running through your head. Perhaps these
few examples have triggered some of your own thoughts, as you are no longer
reading this article: your eyes are just glazing over these words. If this just
happened and you rejoin us a few sentences below, you have just experienced
an intrusive thought.
We all seem to have an intuitive understanding of what an intrusion is.
Yet, attempts to define it immediately sets off debate. According to the one
textbook that has been written on this is, the classic definition of an intru-
sive thought is “any distinct, identifiable cognitive event that is unwanted,
unintended, and recurrent. It interrupts the flow of thought, interferes in task
performance, is associated with negative affect, and is difficult to control”
(Clark 2005:4). Central to this definition is the assumption that an intru-
sive thought always constitutes an unpleasant experience which negatively
impacts functioning. Here, we revisit the definition of intrusive thinking,
by considering all the circumstances in which intrusions might occur, their
manifestations across health and disorders, and their neurocognitive basis.
We start with a rather narrow, presumably more commonly accepted defini-
tion of intrusions being conscious, involuntary, unwanted thoughts, and ar-
rive at an alternative, more inclusive definition of intrusions as interruptive,
salient, experienced mental events. We discuss current methodologies used
to study intrusive thinking, highlighting existing strengths as well as areas
where more research and/or technical innovation are needed. We conclude
with a discussion of the theoretical, therapeutic, and societal implications of
intrusive thinking and its control.
What Are the Everyday Manifestations of

Intrusions and Their Control?
Although intrusive thoughts are often associated with mental health disor-
ders, they also occur to healthy individuals in everyday life (Purdon and Clark
1993; Berntsen 1996). Identifying broad circumstances under which these
thoughts occur to people in general, irrespective of mental health status, may
highlight their adaptive functions in healthy individuals and illuminate the
processes that generate such intrusions. Moreover, considering the motiva-
tions that people may have for controlling intrusions in daily life can shed
light on important psychological and social functions that the capacity for
control helps to support.
Neuropsychological Mechanisms of Intrusive Thinking 127
Types of Content That Drive Recurring Intrusive Thoughts in

Healthy Individuals
Although unintended thoughts can occur for any content, not all such thoughts
are intrusive and bothersome. There are, however, certain contents that consis-
tently trigger intrusive thoughts in otherwise healthy people. Here we discuss
several examples and consider why this content is intrusive.
Emotionally Salient Events. Events that trigger intense emotions, such as

psychological trauma, often lead to intrusive reminders, and the period of their
intrusiveness extends for varying durations, depending on the event. Emotional
events that trigger unwelcome intrusions are typically negative (e.g., anger,
guilt, shame, sadness, fear, and embarrassment). However, positive events are
also capable of triggering repetitive thoughts about pleasant memories, and this
can be quite disruptive, especially when it interferes with focused attention that
people need to perform certain tasks or activities (e.g., job responsibilities).
Incompletions. When people initiate a process and are then unable to com-
plete it, due to an interruption or some other impasse, thoughts related to that
incompletion tend to recur until the process is completed (Horowitz 1975).
This hypothesis is reflected in the classical Zeigarnik effect proposed in the
early twentieth century by Gestalt psychologists, who posited that people had
superior memory for interrupted processes (Zeigarnik 1938). Incompletions
can include both physical and mental tasks, as long as there is an unresolved
problem. An interesting possibility is that the tendency for incompletions to
precipitate intrusive thoughts may be amplified by the salience or emotional
intensity of the incomplete process (Horowitz 1975). Intrusive thoughts related
to the interrupted process (i.e., seeking comprehension, solving a problem)
may persist until the situation is understood. This suggests that emotionally
intense events may not intrude merely due to salience or encoding strength,
but because they are accompanied by a compelling desire to comprehend them.
Relatedly, when people reach a significant impasse in solving a difficult
problem, they continue to “work on the problem” in the background. For ex-
ample, research on creativity and insight problem solving suggests that when
people allow a period of incubation, insights may emerge spontaneously (the
“aha” phenomenon), as though a process has occurred in the background (Gable
et al. 2019). The commonly experienced “tip of the tongue” phenomenon in
memory retrieval (described initially by W. James and S. Freud) is another
example of when a temporarily forgotten item finds its way back into con-
sciousness unpredictably. The content of people’s mind wandering (i.e., when
thoughts distract us from a task at hand) often includes unresolved problems
salient to the individual, whether emotionally significant or not (Smallwood
and Schooler 2006; Klinger and Cox 2011).
Intentions. A special case of “incompletions” worth distinguishing concerns

incompletions arising from intended actions that must be deferred, an ability
studied in research on prospective memory. This research indicates that al-
though people often rely on the environment to remind them to perform inten-
tions, or use intentional self-reminding strategies (e.g., lists), thoughts of the
intended action can still pop into a person’s mind, unbidden, particularly when
intended actions have high importance or affect. Such cases suggest that inten-
tions are maintained in an elevated state of accessibility, even without con-
scious rehearsal. Consistent with this possibility, deferred intentions often have
a special active status in memory (outside of intentional rehearsal) that disrupts
ongoing task performance, as illustrated by intention interference (Goschke
and Kuhl 1993; Cohen et al. 2011; Bugg and Streeper 2019).
Anticipated Events. Although intrusive thoughts often involve past events

or general ideas, they also concern events that have yet to happen. The as-
sociated anticipation can carry emotionally positive, negative, or even both
consequences. Under some circumstances, anticipated events elicit conflicting
feelings of excitement and apprehension in parallel. For example, an upcoming
biopsy may return a diagnosis of cancer or good health.
Uncertain Events. When people are uncertain about a past or future event,
this can promote intrusive thoughts for several reasons (Grupe and Nitschke
2013). For example, if someone is uncertain about whether they have already
performed an action (e.g., locking the front door or taking one’s pills), associ-
ated uncertainty may precipitate worry. Uncertainty about future outcomes
can also result in persistently intruding thoughts, which in turn may prepare
people for different outcomes, thus enabling them to be ready to respond
appropriately.
Dissonant Facts, Events, or Beliefs. When a new fact, experience, thought,

or impulse conflicts with one’s beliefs or self-image, the resulting dissonance
creates a tension that must be resolved. For instance, if a person commits an
action that conflicts with their self-image, the implications for how they should
revise their self-perceptions can be distressing. Are they the good person they
think they are or not? Dissonance creates conflict that can trigger recurring au-
tomatic thoughts until the conflict is resolved. The desire to resolve the mental
discomfort and stress induced by such discrepancies suggests that people strive
for psychological consistency, an idea first proposed by Leon Festinger (1962)
in his theory of cognitive dissonance.
Frequent Events, Stimuli, or Ideas. Not every intruding thought is emotional

or related to incompletion or dissonance. Sometimes, the frequency of an
event, image, or thought induces further repetitions. One striking example is
the repeated hearing, in our imagination, of a song that we have recently heard
on numerous occasions, known colloquially as “ear worms” (Hyman et al.
2013, 2015; Moeck et al. 2018). Research suggests that whether an ear worm
develops is related to how often the song has recurred in recent experience.
Images. As the foregoing examples illustrate, intrusive thoughts can occur

in many representational formats. In some cases, intrusive thoughts can be
verbal or propositional in nature, as might occur in persisting thoughts about
incomplete processes, unsolved problems, deferred intentions, or incongruous,
dissonant occurrences. In other cases, intrusive thoughts can be more sensory
in character, including intrusions from any sensory modality. These mental
images may play a special role in intrusive thinking, a possibility suggested
by their prevalence in psychiatric disorders. Mental images are experiences
of perception that occur in the absence of external sensory input (Kosslyn et
al. 2006; Pearson et al. 2015). They are not limited to remembering the past,
but can also include imagined future scenarios. While often benign, highly
emotional images (e.g., of an upsetting event) can set off a cascade of other
disruptive cognitive and emotional processes, including increased physiologi-
cal responses and rumination (Lang 1977; Grey and Holmes 2008; Holmes
and Mathews 2010; Ji et al. 2016; Holmes et al. 2017). Later, in discussions
of clinical manifestations of intrusive thoughts, we revisit intrusive imagery in
diverse psychiatric disorders.
External and Internal Factors Triggering Intrusive

Thoughts in Healthy Individuals
The previous discussion reviews content that is especially prone to generate in-
trusive thoughts, but it is worth separately considering the conditions that gen-
erate intrusions. In general, the likelihood of intrusions occurring in a given
moment is related to both the presence of retrieval cues as well as the availability
of control resources to resist unwanted intrusions. Here we discuss the critical
role of cue-driven retrieval, matching physiological and mood state, monitoring
processes that may cue thoughts, and the availability of inhibitory control.
Cue-Driven Retrieval. Intrusive thoughts are often triggered by associations

to environmental cues (Berntsen 1996). The most obvious example arises
when a stimulus elicits an unwelcome reminder of a past event. For instance,
after an argument, a friend’s face may remind someone of the altercation, or
a song may bring back memories of a loved one who passed away. Although
intrusive memories provide clear cases of cue-driven retrieval, other forms
of intrusive thoughts, such as feared future events, or unpleasant ideas or im-
ages, also seem to be cue driven. When the same content intrudes more than
once, the content is reemerging from memory of previous thoughts. One can
frequently identify specific cues in the environment, or in a person’s patterns
of thinking, that drive retrieval of perseverative content (e.g., Gagnepain et al.
2014; Benoit et al. 2016).
Matching Mood and Physiological State. Sometimes the cues that trig-
ger retrieval are not specific stimuli in the world or concepts in memory, but
rather are broad psychological or physiological states (Berntsen 1996). It is
well established that information and experiences are often stored in memory
in association with a representation of the context in which they took place
(for a review, see Anderson and Hanslmayr 2014). This can include not only
the spatiotemporal context (i.e., the environment), but also the mood, drug, or
arousal state present at encoding. Later on, the chance of retrieving the content
is higher when a person’s state resembles the one at encoding. For example,
experiences encoded in a sad or angry mood are more likely to be recalled at a
later time, when a person is once again in a sad or angry mood than when they
are in a happy mood.
Diminished Cognitive Control. In addition to the particular cues at retrieval

and their match to encoding, other elements of the state of the person at retrieval
may influence the frequency and persistence of intrusive thoughts. An impor-
tant category of state-related variables that can influence intrusion frequency
is whether a person is suffering from diminished cognitive control that might
otherwise help the person to limit involuntary retrievals. If mechanisms such as
inhibitory control prevent or reduce intrusions, then anything that compromises
these abilities is a risk factor for intrusiveness, even in healthy individuals.
Several factors common in healthy samples may give rise to such deficits in
cognitive control, including sleep deprivation (Nilsson et al. 2005; Drummond
et al. 2006), general fatigue, stress (Shields et al. 2016), lack of exercise
(Hillman et al. 2008), and intoxication. Although such changes in state are
temporary for most healthy individuals, chronic depletion of cognitive control,
especially with several of the above factors contributing, could put healthy
individuals on a trajectory to develop persistent intrusions and changes in
mood that are clinically significant. In addition, consistent, trait-related defi-
cits in cognitive control may give rise to significant risks in controlling intru-
sive memories and thoughts, and may be a risk factor for psychiatric disorders
(Levy and Anderson 2008).
Desirability of Control
In healthy individuals, when intrusive thoughts occur, they are usually per-
ceived as unwanted, at least at that moment. As a result, people often try to
exclude them from awareness in an effort to regain control over thoughts and
emotions. When successful, such efforts enable a person to put unwelcome
thoughts out of mind, thus diminishing their accessibility in memory and re-
ducing their tendency to return. Such attempts to facilitate the forgetting pro-
cess often serve important behavioral, emotional, and social functions. Here,
we consider several contexts in everyday life in which people are motivated to
forget thoughts for a functional reason.
Concentration during Tasks. Sustained focus on a necessary task can be dif-

ficult when unrelated distracting thoughts intrude into awareness (Smallwood
and Schooler 2006). Such thoughts can be of any valence. For example, receiv-
ing good news or bad news prior to leaving for work can undermine our focus.
If distracting thoughts slow progress or increase the propensity for errors, it
becomes essential and adaptive to put the distracting thoughts out of mind.
Concentration is an extremely common selective attentional mechanism for
trying to control intrusive thoughts, and successfully achieving concentration
is clearly adaptive.
Executing High-Performance Cognitive and Motor Skills. Professional ath-

letes are often under extreme pressure to perform to a very high standard, and
when they do not, there can be significant consequences for themselves or their
teammates. In sports psychology, a literature on the causes of “choking” under
pressure (Beilock and Gonso 2008) has emerged to address the processes that
lead an athlete’s performance to deteriorate under pressure. This often includes
the inability to overcome intrusive thoughts that undermine the focus needed
for top performance.
Regulating Pain. Thoughts about pain, what it feels like, how uncomfortable
it is, and worries about what it may mean often intrude persistently, even after
the pain has resided. We consider people who can manage intrusive thoughts
about their pain to be resilient and able to cope. In contrast, those who are over-
whelmed by pain, who allow their discomforts to blossom into catastrophic
thinking, not only pay a price for this distraction, their suffering becomes mag-
nified and extended (Edwards et al. 2006). When confronted with a painful
stimulus (e.g., the cold pressor task), people with a moderate history of adver-
sity tolerate the pain longer and show less pain catastrophizing (the label given
to excessive thoughts about the pain), relative to people who have no history
of trauma (Seery et al. 2013). It appears, therefore, that in the face of physical
discomfort, resilience emanates from the capacity to control intrusive thoughts
about the painful stimulus, which not only reduces suffering but increases the
ability to pursue other goals.
Regulating Affect. When unwelcome thoughts intrude, they often evoke un-
welcome changes in emotional state. For example, being reminded of an ar-
gument may trigger anger; images of an upcoming doctor’s visit may trigger
fear or anxiety; or seeing the same car that your ex-partner used to drive may
evoke sadness. As a result, intruding thoughts trigger mechanisms that regu-
late emotion to return a person to a neutral or positive state. Although these
endogenously triggered emotions are common and contribute substantially to
psychological disorders, they have been neglected in research on emotion reg-
ulation, which has focused on regulating affect triggered by external emotion-
eliciting stimuli. Understanding how intrusive thoughts can be downregulated
in memory may provide critical insights into how people regulate the emotions
that they elicit, contributing to successful emotional homeostasis (Engen and
Anderson 2018).
Persisting in the Face of Failure. When memories or thoughts of past failures

dominate someone’s thoughts, they can be disruptive. Intrusive thoughts about
failure are unpleasant and undermine feelings of competence and control; this
may lead a person to abandon their goals earlier than they should or to fail to
improve their performance. When someone can set aside past failures, this en-
ables them to improve their skills or knowledge or to find creative solutions to
the problems that led them to fail. Persistence, especially dogged persistence in
the face of challenges and setbacks, requires successful regulation of thoughts.
If persistence is an adaptive trait enabling major personal achievements, the
inability to control thoughts of failure can limit personal growth.
Protecting Self-Image. People sometimes feel embarrassed or ashamed

about things that they have done. Alternatively, others may say hurtful things
that undermine self-confidence. Although people vary in how they respond
to such events, they often try to put the unwelcome content out of mind.
Correspondingly, they show greater forgetting for negative feedback, exhibit-
ing a remarkable capacity to not remember their faults or misdeeds. Work on
mnemic neglect (Sedikides and Green 2009; Sedikides et al. 2016) has es-
tablished this pattern of forgetting and linked it to threats to a person’s self-
concept: recent threats to one’s self-image are forgotten. Moreover, work on
the positive self-illusion (Taylor and Brown 1988) has linked this tendency to
improved mental and physical health: healthy individuals have a higher view
of their capabilities than would be supported by observers. In contrast, de-
pressed individuals, ironically, often have a more accurate view of their cir-
cumstances. Similarly, most individuals show a very powerful and replicable
positivity bias in terms of the autobiographical memories that they ultimately
retain (Walker et al. 2003), whereas depressed individuals very often show the
reverse tendency.
Justifying Inappropriate Behavior. Sometimes people commit unethical, im-

moral, or hurtful acts about which they feel shame. Despite this, most people
like to think of themselves as decent, creating dissonance between one’s be-
liefs and deeds. Intrusive thoughts about these discrepancies often develop and
must be addressed. Research on ethical amnesia indicates that people often
forget their ethical lapses (Kouchaki and Gino 2016; Stanley and De Brigard
2019), suggesting that people suppress these uncomfortable thoughts.
Maintaining Attitudes and Beliefs. A person’s political, religious, or per-

sonal beliefs are often resistant to contradictory evidence. When new facts
contradict our beliefs, dissonance is created that must be resolved. This reso-
lution often involves controlling intrusive thoughts to forget the inconvenient
information. For example, Republicans and Democrats show enhanced di-

rected forgetting for attitude statements that are incongruent with their be-
liefs, compared with congruent statements (Waldum and Sahakyan 2012).
Moreover, a person’s memory can be shaped by selectively recounting an
event (Cuc et al. 2007; Stone et al. 2012, 2013), a form of thought substitu-
tion (Benoit and Anderson 2012; Anderson 2001). Intriguingly, this can un-
dermine memories of omitted facts, a phenomenon known as socially shared
retrieval-induced forgetting (Cuc et al. 2007; Stone et al. 2012, 2013). The
selective forgetting that people exhibit for facts that are incompatible with
their beliefs suggests that healthy individuals control their thoughts to avoid
the discomfort of inconsistency.
Forgiving Others and Maintaining Attachment. Sometimes friends or rela-

tionship partners commit offenses that provoke anger, giving rise to persistent
intrusive thoughts. Indeed, intrusive thoughts about anger can be intense for
some people and can take considerable time to “get over,” placing strain on per-
sonal relationships and requiring the ability to override thoughts and impulses
originating from the anger. This ability is predicted by individual differences in
inhibitory control, suggesting that overcoming intrusive thoughts involves, in
part, control processes that suppress them in service of forgiveness. Consistent
with this possibility, when people decide that an offense may be forgiven, it is
easier to suppress intrusive thoughts about the offense, ultimately leading to
worse memory (Noreen et al. 2014). Thus, suppressing intrusive thoughts of
anger ultimately contributes to a healthy capacity to forgive and forget in social
relationships.
Similar considerations apply when there is a need to maintain an attach-
ment relationship with a parent, guardian, or powerful authority figure (e.g., a
boss), which may be essential to survive or thrive in an environment. In such
cases, the capacity to control recurring intrusive thoughts enables a person to
maintain normal relationships despite this conflict and, in some cases, to forget
about the discrepant content.
Mind Wandering: An Everyday Manifestation of Intrusive Thinking?
One interesting example of everyday intrusive thinking is the widely studied

experience of mind wandering. Returning to our example at the beginning
of this chapter, we are all familiar with suddenly realizing that while our
eyes have been moving across the page, our minds have been temporarily
sidetracked by thoughts unrelated to the text. Such a situation is known as
mind wandering (Smallwood and Schooler 2006). Using thought sampling
techniques in which individuals are intermittently queried as to whether
their thoughts are engaged in the task at hand, research suggests that people
mind wander up to fifty percent of their waking hours (Killingsworth and
Gilbert 2010). The surprising frequency with which people engage in mind
wandering raises an intriguing question: Should mind wandering be viewed

as everyday intrusive thinking? Let us consider the following aspects of this
question.
Unwanted. Mind wandering is often unwanted because it disrupts task per-

formance across many cognitively demanding domains (for a review, see
Smallwood and Schooler 2015). It is, for instance, a major source of car ac-
cidents (Gil-Jardiné et al. 2017). Yet while mind wandering interferes with
current tasks, it can be useful for completing more distal goals, such as plan-
ning (Baird et al. 2011) and creative problem solving. For example, when a
vigilance task was interposed between two trials of a creativity task, mind
wandering interfered with the vigilance task but enhanced performance on
the second round of the creativity task (Baird et al. 2012). In a diary study,
both creative writers and physicists indicated that twenty percent of their
ideas occurred outside of work (Gable et al. 2019). Thus, although mind
wandering can interfere with the task at hand, it often contributes to progress
on other problems.
Unintended. Individuals frequently report mind wandering and are often not
even aware that they are engaged in it, until they are caught, as in the expe-
rience sampling probe (Schooler et al. 2011). Nevertheless, individuals may
sometimes deliberately abandon a task in favor of other thoughts (Seli et al.
2015). The lack of awareness that one is mind wandering (i.e., lack of meta-
awareness; Schooler 2002) contributes to more disruptive mind wandering
episodes (Schooler et al. 2011), which are neurologically distinct from on-task
thinking (Christoff et al. 2009). Notably, lack of meta-awareness has been as-
sociated with mind wandering about unwanted thoughts (Baird et al. 2013a).
This suggests that intrusive thoughts may similarly slip below the radar of
meta-awareness (Takarangi et al. 2014).
Recurrent. Although up to fifty percent of people’s waking hours is spent

mind wandering (making mind wandering a recurring form of thought), the
content of mind wandering routinely varies from one episode to the next.
Hence, whether mind wandering is recurrent depends on the level of analysis.
Associated with Negative Affect. In their influential paper, Killingsworth and

Gilbert (2010) hold that “a wandering mind is an unhappy mind.” This charac-
terization reflects the observation that when individuals mind wander, they are
routinely less happy than when they are on task. Nevertheless, the degree to
which episodes of mind wandering drive negative affect may be partly deter-
mined by content. For example, Franklin et al. (2013) found that participants
were generally less happy when mind wandering than when on task, unless
they were mind wandering about something interesting, in which case they
were actually happier.
Difficulty of Control. Individuals’ perceptions regarding the controllability of

mind wandering varies considerably (Zedelius and Schooler 2017), and these
perceptions influence its frequency and impact. Individuals who view mind
wandering as outside of their control mind wander more often, and their per-
formance is more disrupted than those who view it as a mental state they can
manage. Collectively, evaluation of the relationship between mind wandering
and the definition of intrusive thoughts indicates that mind wandering is often,
but not always, a form of everyday intrusive thought.
Summary and Commentary
As the above discussion illustrates, intrusive thinking occurs in a range of

circumstances in healthy individuals. Such thoughts do not, by themselves,
indicate a mental health disorder but are a normal feature of life. Moreover,
attempts to suppress intrusive thoughts are often not only highly successful
in healthy individuals, they are in fact essential to achieving numerous social,
emotional, and cognitive goals, including the successful regulation of emo-
tional state (Engen and Anderson 2018). Not all healthy people are equally
effective at controlling intrusive thoughts, however, and vulnerability to per-
sistent intrusions is critical to understand. These observations highlight the
importance of understanding when and how intrusive thoughts become patho-
logical and contribute to psychological disorders.
What Are the Main Manifestations of Intrusive

Thinking in Mental Health Disorders?
As indicated in the preceding sections, occasional intrusive thoughts are nor-

mal occurrences and even commonplace in certain circumstances. However,
there are a number of mental health disorders for which intrusive thinking is
a core symptom, becomes problematic, and can interfere with function. Some
common disorders in which intrusive thinking is a critical element are briefly
described below; for further detail, see Schlagenhauf et al., Banich, as well as
Brewer et al. (this volume).
Symptoms of intrusions in posttraumatic stress disorder (PTSD) include recur-

rent, unwanted, involuntary, and distressing memories of a traumatic event
(intrusive memories); reliving the traumatic event viscerally, as if it were hap-
pening again (flashbacks); upsetting dreams or nightmares about a traumatic
event; and severe emotional distress or physical reactions to cues that remind
an individual of the event (e.g., the smell of burned rubber reminding someone
of a car accident or a ski mask that triggers memory of an armed robbery).
Intrusive memories or flashbacks are often precipitated by environmental cues

that remind an individual of the traumatic event. Flashbacks are immersive,
causing individuals to “dissociate” from their bodies and feel that they are
back in the situation in which the trauma occurred. It is not clear whether
intrusions and flashbacks are qualitatively different phenomena or fall on a
spectrum of severity in which flashbacks represent a severe form of intrusions.
The intensity and content may be different, yet both are multimodal sensory
images that are interruptive, salient, experienced, unwanted, involuntary, and
often recurrent.
Obsessive-Compulsive Disorder and Related Disorders
In obsessive-compulsive disorder (OCD), intrusive means the feeling of “be-

ing out of control.” This feeling is experienced in different phenomenological
domains since OCD develops over time. OCD is a process with different clini-
cal stages rather than one single stage. These stages follow a dialectic interac-
tion in which an intrusive event elicits a response and the response amplifies
the intrusive event. Through different neurobiological adaptations, the course
of OCD eventually worsens. Thus, OCD should be regarded as a disease pro-
cess that develops through the amplifying interaction between (the reflection
and resistance of) the person (mind) and the disorder (brain). The following
example describes how intrusions may develop in OCD.
A young mother recently gave birth for the first time and now carries the
responsibility of brand-new motherhood. Her partner leaves daily for work,
and thus she is home alone with the child. Viewing her young baby in the crib,
a thought appears in her mind: she imagines that she could strangle her baby
in the crib and that no one is there to prevent her. The mere presence of the
thought is intrusive because it occurs beyond her free will. She feels out of
control and is unable to control her thinking. She is worried because the idea
of strangling her baby does not fit her ideal of motherhood. The content of the
thought is intrusive and ego-dystonic because it is not in line with her self-
image. She wonders whether she could really strangle her baby and begins to
feel out of control. How could she be certain that she would not strangle her
baby, given the fact that humans are notoriously unpredictable? As the implica-
tion of the thought is intrusive, she feels anxious because the thought confronts
her with being out of control.
The emotional value (anxiety) of the thought is intrusive. The presence,
content, implication, and emotional value of the thought all have an intru-
sive quality. She actively resists the thought because it annoys her and feels
intrusive. The process of reflecting or resisting, however, serves to reinforce
the frequency and intrusive strength of the thought. The thought becomes
obsessional. Her attention is completely drawn to that one single thought.
Obsessionality is a dysfunction of intentionality: the incapacity to shift focus
or attention to another topic due to a stronger and longer intentional relation
with the mental act. The thought is intrusive because of its obsessive nature.
She cannot suppress the thought. Moreover, she is compelled to think about
her obsession. Compulsivity is a dysfunction of sense of agency: she is forced
to think about the intrusion, contrary to her willpower. The thought is intru-
sive because of its compulsive nature. Gradually the thought becomes more
present and repetitive; it loses its original meaning, but becomes intrusive
because of its duration and repetition. The thought has become a full-blown
obsession. Obsessions are answered with compulsions (e.g., obsessing about
germs could lead to compulsive handwashing). Though initially successful
in reducing anxiety, gradually they become intrusive since the acts have to
be performed compulsively. Note that both obsessions and compulsions are
intrusive, and both have an obsessional and compulsive quality. Eventually,
the anticipatory power of the intrusion is so overwhelming that reality test-
ing gets disturbed. She does not know anymore whether she has or has not
strangled her baby. Thoughts may take on a delusion-like character, with psy-
chotic features.
Substance Use Disorder
Intrusive thinking in individuals with substance use disorders (SUDs) may

include planning to procure the drug, recall of the experience related to its
use, as well as the anxiety related to lack of access to or possession of the
drug. However, there is large variability between patients, dependent on dis-
ease stage. In the initial stages, the individual may perceive thoughts of the
drug as non-intrusive and innocuous or even pleasurable, at least as long as
they believe to be in control of drug intake, and that such thoughts do not sig-
nificantly interfere with daily activities. Once the disease progresses toward
more severe stages, the individual begins to perceive thoughts of the drug as
intrusions. This usually happens when the individual realizes that they have
lost the ability to make decisions regarding whether or not to obtain and/or
take the drug.
In individuals with SUD, the onset of intrusive thinking varies widely.
Intrusive thoughts may appear during acute withdrawal from the drug, or years
after the last exposure to the drug. Furthermore, they can be elicited by an in-
ternal cue (e.g., a memory of an event related to the drug, or a particularly emo-
tionally negative moment while the patient is alone at home) or external cues,
such as other people taking the drug, movie scenes, or physical cues around
the patient (e.g., restaurants, bars, or images related to the drug or addiction).
In conclusion, intrusive thinking is a major component of SUD as well as
in other addictions, such as pathological gambling. It occurs throughout every
step of the life cycle of SUDs and, in some cases, can be the primus movens of
a series of actions and events leading to relapse and drug taking.
Mood Disorders
Clinically, intrusive thoughts are neither a core nor a necessary feature of mood
disorders. However, they appear regularly in a manner that has important bear-
ing on its treatment, severity, and outcome, specifically in the form of rumina-
tion, suicidal thoughts, negative automatic thoughts and “flight of ideas.”
Rumination is a frequently observed cognitive feature of depression, in-
volving the repetitive and persistent focusing on the causes of the current
state of distress and its likely consequences. Depressive rumination is typi-
cally centered around personal shortcomings, faults, failings, and mistakes
(Treynor et al. 2003). Rumination has intrusive features: it is highly interrup-
tive and distracts individuals from engaging in other tasks. While individuals
often state they do not want to ruminate, they also hold metacognitive beliefs
that it is important to do so (Borkovec and Roemer 1995). Rumination is asso-
ciated with the onset of depression and, when combined with negative cogni-
tive styles, predicts the duration of depressive symptoms (Nolen-Hoeksema et
al. 2008). The extent to which it is hence desired and indeed under volitional
control is unclear.
Individuals suffering from suicidal thoughts report at times imagery related
to potential ways of committing suicide (Holmes et al. 2007). This imagery can
be experienced as unwanted, intrusive, and interruptive.
Negative automatic thoughts are regarded as a central feature of depres-
sion in cognitive models of depression (Beck 1976). In this view, events that
are related in some form to core negative beliefs or schemata can trigger fast
interpretations or “automatic thoughts,” which, due to their negative nature,
promote a depressed, negative mood. These automatic thoughts appear very
rapidly and can profoundly influence behavior. However, as these automatic
thoughts closely relate to core beliefs, they tend not to be experienced as intru-
sive or unwanted.
Individuals with manic episodes of bipolar disorder can display a “flight
of ideas,” whereby they exhibit a rapid sequence of unrelated thoughts. This
state is often experienced in the context of mania and tends to be positively
experienced as a phase of heightened creativity. In mixed states, however, it
can also be perceived negatively and exhibit loss of control of one’s thinking.
Hence, although intrusive thoughts are not central to mood disorders, they do
feature prominently and in areas that are thought to be closely related to the
core mechanisms of the illnesses.
Anxiety
Two subtypes of anxiety have been distinguished: anxious arousal (e.g., panic)
and anxious apprehension (e.g., worry). In cases of anxious arousal, the in-
dividual has fearful reactions and thoughts that lead to somatic symptoms
and/or somatic symptoms are interpreted in a fearful manner. For example,
in simple phobias, an individual will see a phobic object (e.g., a spider) and
have intrusive thoughts regarding that object (e.g., “The spider is about to
crawl on me and bite me”) that are often accompanied by somatic symptoms
(e.g., sweating, increased heart rate). In cases of panic disorder, the individual
has somatic symptoms, such as increased heart rate and dizziness, associated
with intrusive thoughts (e.g., “I am having a heart attack and am going to
die”). This situation of a bodily state leading to intrusive thoughts is some-
what akin to SUDs, in which a somatic state (i.e., withdrawal) can lead to
intrusive thoughts of craving.
In contrast, in cases of anxious apprehension, intrusive thoughts tend to
be related to a future event: about the feared event itself, ways to avoid the
fear, or the discomfort/harm that might be associated with the event. In some
cases, intrusive thoughts can be more distressful when a person anticipates
the event, than during the event itself. This appears to be associated with
the fact that individuals with anxious apprehension have an intolerance of
uncertainty (e.g., not being 100% sure that the flight will not involve severe
turbulence).
Individuals that experience anxious arousal and/or anxious apprehension
may know that their fears are likely unrealistic or overblown. This knowledge,
occurring simultaneously with the experienced fear and intrusive thought,
can cause distress and lead to additional intrusive thoughts related to poor
self-evaluation similar to depressive rumination: “What is so wrong with me
that I cannot get on an airplane like everyone else? Why am I such a baby?”
However, intrusive thoughts associated with anxiety tend to be more future
oriented whereas intrusive thoughts associated with depression tend to be more
concerned with past events.
Psychosis
Psychosis is a state characterized by hallucinations and delusions, which is

common in schizophrenia spectrum disorders but also observed transdiagnos-
tically. Hallucinations are false perceptions, mostly experienced with simi-
lar sensory quality, as if originating from an outside stimulus but without an
external stimulus being present. For example, a patient with schizophrenia
might repeatedly experience hearing voices that are negatively commenting on
what they are doing and perceive the voices to come from outside their head.
Auditory verbal hallucinations of this type can repeatedly occur and severely
interfere with daily life, although frequency and distress vary.
Delusions are false beliefs held with high subjective certainty and confi-
dence despite contrary evidence. Delusional ideation is very salient for the
patient and center on the person. For example, common topics are delusions
of reference or prosecution. Deluded persons do not experience the delusional
ideation as interruptive to one’s thoughts. Delusional ideations are part of the
person’s belief system and thus are ego-syntonic; still, they can be disruptive
for the person’s relation to their interpersonal and social surrounding. A special
case is delusion of control such as thought insertions, when (mostly verbal)
thoughts are experienced as being “inserted” by another agent into one’s head.
Inserted thoughts can be salient, interruptive to the train of thought, and un-
wanted, but are attributed to an outside agent rather than to oneself, and there-
fore cannot be experienced as ego-syntonic nor ego-dystonic.
Attention Deficit Hyperactivity Disorder
Attention deficit hyperactivity disorder (ADHD) is a neurodevelopmental dis-

order characterized by a persistent pattern of inattention and/or hyperactivity-
impulsivity that interferes with everyday functioning (American Psychiatric
Association 2013). People with ADHD are highly distracted by things happen-
ing in the outside world as well as by internal thoughts. The issue is whether
these distracting internal thoughts should be classified as intrusions. To an
external observer, they are certainly disruptive and maladaptive. Compared
to other disorders, however, the intrusions may, on average, be much less sa-
lient since an individual with ADHD can be distracted by any event, salient
or trivial. The distraction can be relatively diffuse, and thoughts may not have
specific, recurrent content. While certainly maladaptive and poorly controlled,
these thoughts are also not necessarily unwanted; they may function as a wel-
come distraction (e.g., when feeling bored). An experience sampling study
found that individuals with ADHD symptoms experienced excessive disrup-
tive mind wandering, together with little meta-awareness on how to regulate
this (Franklin et al. 2017). In general, one could ask whether having an ex-
ceptionally low threshold for internal and external distraction might render
irrelevant the concept of intrusions; in such cases, the inability to focus and fol-
low goals (i.e., stay on task) changes the goal hierarchy. Thought meandering
becomes the default; there is simply no focused thought process upon which
this meandering intrudes. This lower threshold could make people with ADHD
more vulnerable to other disorders and the experience of intrusions, in general
(Abramovitch and Schweiger 2009).
Revisiting the Definition of Intrusive Thinking: A Synthesis
From the description of all the possible manifestations of intrusive thinking in

everyday life, as well as across clinical syndromes, a number of features can
be distilled that best capture what intrusions are. Each of these features will be
discussed below. From this, we select what we think are the key features of in-
trusive thinking to arrive at two definitions: one narrow, probably more typical
definition of intrusions, and a broader, more inclusive one.
Consciousness. To what degree are intrusions conscious? Is it possible to

have unconscious intrusions? It is hard to conceive of an example where an
intrusion is unconscious. However, the term consciousness is often confused
with meta-consciousness (Schooler 2002); that is, being aware that something
is an intrusion rather than merely experiencing it. Because of this ambiguity,
experience may be a less controversial term, though arguably more abstract.
In addition, while intrusions can encompass a range of unexpected events or
actions, intrusive thinking explicitly refers to cognition. This may be incom-
patible with nonhuman animal models and may also overestimate how vol-
untary our usual thoughts are (see also Liu and Lau, this volume). Instead of
conscious thought, we suggest the broader term experienced mental events.
Unwantedness/desirability. Although intrusions are not necessarily negative

in content, they may be unwanted in the moment, as they distract from the task
at hand. The question then is: Does it matter what the task at hand is? Are in-
trusions during daydreaming unwanted, as long as they have no clear negative
content? Can something even be identified as an intrusive thought if it occurs
during daydreaming? One can even think of examples of clear pathological
intrusions that are not necessarily unwanted. In depression, for instance, rumi-
nation is experienced as a strategy to solve a problem. In some cases, it reflects
long-term aspirations or a society that makes something unwanted. It boils
down to the question of what wanting is; different individuals (or even differ-
ent voices within an individual) interpret what is wanted differently in the mo-
ment. Because unwantedness may not be a defining feature in every instance of
intrusive thinking, it seems necessary to develop a narrower clinical definition
that is more closely related to the patient’s experience of being out of control.
Involuntary/Controllability. Generally we think of intrusions as being invol-

untary, though this can mean different things. Involuntary can refer to the fact
that a thought was unintended or uncontrollable once it appears. Are thoughts
that are unbidden, but easily dismissed, intrusive? We suggest including in-
voluntary in a narrow definition of intrusive thoughts, but not in a broader one
(see below).
Disruptiveness. While unwantedness and controllability may not be a defin-

ing feature of every instance of intrusive thinking, a necessary characteristic
is that intrusive mental events interrupt and disrupt current cognition. By defi-
nition, they intrude upon ongoing processes (e.g., a task or a gentle, natural
cadence of unconstrained thought). The disruptiveness may, however, not be
experienced as such by the person with intrusive thoughts, though clearly it has
maladaptive consequences to an external critic.
Salience. Something can only interrupt if it has gravitational pull; that is, if it
captures attention (see also Fedota and Stein, this volume). Such gravitational
pull may lie in different, not mutually exclusive, aspects (e.g., its valence,
highly negative or positive), its vividness, its novelty, or its incongruence with
a current state.
Valence. Inherent to Clark’s definition (Clark 2005:4) is the assumption that

intrusive thoughts are unpleasant, but this is not always the case. Examples of
potentially positive intrusive thoughts may include being in love, or experienc-
ing feelings of grandeur (as occurs during manic or psychotic episodes), or
when one has an insight (“aha” moments). Therefore, we suggest character-
izing intrusive events as salient to the person, but not necessarily as negative
in valence.
Content and Shape. Intrusions can take different shapes: they can present as
verbal thoughts, slips of action, or mental images. Content varies in valence
as well as in time frame, ranging from the past, present, or future. We suggest
not including content or shape as a defining feature of intrusive mental events
as it might vary widely between different instantiations of intrusive thinking in
healthy and clinical states.
Punctate versus Extended. Intrusions usually appear unexpectedly, with a

sudden onset and a limited duration. They can, however, last for an extended
period of time, as in the case of an uncontrolled flashback episodes in PTSD,
or if we include delusions, mind wandering, worry, craving, and rumination as
forms of intrusive thinking. In these latter cases, one could argue that only the
brief, initial episode is the intrusion that sets in motion a cascade of secondary
cognitive processes. Intrusive suggests that something intrudes upon some-
thing; if these secondary cognitive processes continue for an extended period
of time, they become the primary cognitive process, perhaps even changing
the goal hierarchy. Should such a primary cognitive process be labeled intru-
sive thinking?
Recurrence. While an ongoing process should perhaps not be labeled

intrusive thinking, intrusions can repeat themselves and still qualify as
intrusions. In fact, a key feature of intrusive memories in PTSD and other
clinical syndromes is their recurrent nature. It may be the frequency as well
as the recurrent content that distinguishes clinical intrusions from non-
clinical intrusions.
Trigger. As described at the outset, intrusions can be triggered by external cues,

such as certain sights or smells or through internal cues such as mood states.
However, intrusions are more than reflexive orientation to salient stimuli: the
stimulus-evoked mental event has to interrupt the ongoing process significantly.
Agency. As described above, to be defined as intrusive, thoughts must be at-

tributed to oneself rather than an external agency.
In conclusion, a commonly accepted definition of intrusions may be that they

are conscious, involuntary, unwanted thoughts. While easy to understand and
applicable to many cases, this definition may mean different things to differ-
ent people and not encompass the entire spectrum of intrusions. An alterna-
tive, more inclusive definition may be that of intrusions being interruptive,
salient, experienced mental events that generally recur, particularly in clini-
cal syndromes. While the latter definition encompasses the clinical definition
of intrusions, it does not differentiate nonclinical from clinical phenomena.
Clinical intrusive thinking differs from its nonclinical form with regard to fre-
quency, intensity, and maladaptive reappraisal. Intrusive thinking with clinical
significance is often part of a clinical syndrome and results in distress and
disability as a general identifier of mental illness. In Table 9.1, we evaluate to
what degree each of these key features for both definitions is manifested across
symptom clusters; for a slightly different conceptualization of intrusions in
pathology, see Monfils and Buss (this volume).
From Table 9.1, it is clear that symptoms occurring in PTSD and OCD
best exemplify the construct of intrusive thinking. Less prototypical exam-
ples of intrusive thinking are flash-forward events, drug craving, suicidal
ideation, rumination, and worry. In these instances, intrusions may not al-
ways be involuntary, unwanted, or clearly interrupt a task at hand. Rather,
patients may voluntarily engage in these thoughts, for example, as a way to
dampen present agony (suicidal ideation) or to mentally expose oneself to
a feared situation in the future (worry), thereby reducing the risk of being
overtaken by panic unexpectedly. In drug craving, these thoughts are only
unwanted and interruptive to the degree that an individual intends to abstain
from using. It is important to note that while worry is part of the DSM-5 cri-
teria for generalized anxiety disorder and as such is a real symptom, rumina-
tion is considered a risk factor for depression, but not a diagnostic criterion.
However, both constitute repetitive negative thoughts, with common under-
lying features as well as unique features (e.g., Hur et al. 2017), and seem to
qualify as intrusive thinking.
It is less clear to what degree symptoms in ADHD qualify as intrusive
thinking and, relatedly, mind wandering in healthy individuals. Typical daily
thought meandering certainly does not elicit the derailing effects that clinical
intrusions do and may not even be interruptive. Impulsive actions and tics do
not seem to qualify as they are not clear experienced mental events (i.e., they
do not necessarily involve awareness). Finally, it is unclear to what degree
hallucinations, thought insertions, or delusions should be considered examples
of intrusive thinking. These thoughts may not generate the same feeling of
uncontrollability and intrusiveness, and an individual may not even realize that
these are intrusive thoughts, such as when an individual truly believes that
the police or secret service is monitoring their thoughts, or when someone is
getting commands via imagined voices. Hallucinations and thought insertions
Table 9.1 Features of intrusive thinking across symptom clusters.

Thought Mental event
Experienced
Interruptive
Involuntary
Conscious
Unwanted
Salient
Symptom Related disorder Further specifier
Intrusive memories PTSD x x x x x x Recurrent,

past oriented
Flashback PTSD x x x x x x Recurrent,
past oriented
Flash-forward Bipolar disorder x ? ? x x ? Future oriented
Obsessions OCD x x x x x x Recurrent
Compulsions OCD x x x x x x Recurrent
Drug craving SUD x x ? x x ? Depends on
motivation for
abstinence;
recurrent,
future oriented
Pathological worry Anxiety x ? ? x x ? Future oriented
Depressive Depression x ? ? x x ? Past oriented
rumination
Suicidal thought Depression x ? ? x x ? Future oriented
Thought insertions Psychosis x ? ? x x ? No agency,
attributed to
outside agent
Auditory Psychosis x ? ? x x ? No agency,
hallucinations attributed to
outside agent
Delusions Psychosis, depres- x – – x x – No agency
sion, mania
Inattention ADHD x x ? x – x Relatively diffuse
Impulsive action ADHD x x ? – ? x
Mind wandering ADHD x ? ? x ? ?
Tic Tourette syndrome – x x – ? x Recurrent
Urge Tourette syndrome x x ? x x x Recurrent
are usually attributed to an outside agent. In our definition of an “experienced

mental event,” a sense of agency is implied.
How to Best Study the Processes Underlying

Intrusive Thinking and Its Control
In this section, we consider a number of general mechanisms that underlie

the likelihood for a person to experience intrusive thoughts, and how they are
assessed. As suggested above, situations that trigger intrusions all tap into
mechanisms related to the ability to control thoughts, actions, and mnemonic
processes. Most of these processes come into play at the time of (involuntary)
memory retrieval or when habits or compulsions interfere with goal-directed
behavior. Here, we elaborate on what is known about these mechanisms and
the methods to investigate them.
Cognitive Control Processes
Action Control
Goal-directed actions are those that we acquire to bring about a change in the
world that accords with our basic desires. In this view, I perform an action, A,
because I want a certain outcome or goal, O, and believe that A is the way to
achieve O. Action A may start life under goal-directed control; however, with
extended training performance, it can become more automatic and invariant,
elicited by environmental stimuli, instead of to achieve specific consequences,
thus becoming habitual. Both actions and habits are acquired and represented
in parallel rather than serially. Therefore, it is possible for certain errors to
emerge as a consequence of two controllers attempting to control action si-
multaneously: suddenly switching from goal-directed to habitual control can,
for instance, result in the intrusive performance of an action that is adaptive
in another situation, but not in the present situation where it is unwanted (e.g.,
continually turning on the windscreen wipers rather than the blinkers in a
rental car). Conversely, intrusive goal-directed control can disrupt the smooth
operation of a well-learned habit. It is well known, for instance, that suddenly
exerting goal-directed control while playing a complex piece of music on the
piano can result in reduced capacity to perform the piece accurately. The same
holds for well-trained athletes (see Balleine, this volume).
Both actions and habits require top-down control to prevent excessive and
hence maladaptive impulsive behavior. Clues about neural systems support-
ing the regulation of intrusive thoughts come from the field of action control,
capitalizing on putative similarities between inhibition of motor responses and
thought control. In the simple Stop-Signal paradigm, for each trial, subjects
get ready to respond when a Go signal occurs; then, on a minority of trials,
they try to stop in response to a subsequent Stop signal (Lappin and Eriksen
1966; Verbruggen and Logan 2008). Anderson and Green (2001) have long
suggested that stopping a memory or thought from intruding is like stopping a
motor response. When the stopping process fails, an intrusion occurs. Several
studies have provided evidence consistent with this, although it is not yet cer-
tain whether there is a common process, common principle, or common cir-
cuitry (e.g., Morein-Zamir et al. 2010; Depue et al. 2016; Guo et al. 2018;
Castiglione et al. 2019). Assuming there is a commonality between the stop-
ping of a motor response and the termination of a memory or thought, it may be
useful to adopt a distinction from the motor stopping literature between global
suppression and selective suppression (Aron 2011). Stopping a memory or
thought can be achieved by a global systemic suppression mechanism, which
suppresses all memories (Hulbert and Anderson 2018), or through a selective
stopping process in which the target is suppressed by retrieving an alterna-
tive memory (Hertel and Calcaterra 2005; Bergström et al. 2009; Benoit and
Anderson 2012).
Pavlovian-instrumental transfer (PIT) provides another experimental par-
adigm relevant to intrusive thinking and may provide evidence of intrusive
thinking in nonhuman animals as well as humans. In the case of positive PIT,
experimental animals (or humans; Freeman et al. 2014) are trained on two
actions to earn distinct reward outcomes: A1 → O1 and A2 → O2. They are
then exposed to two stimuli paired with the same outcomes but without the
opportunity to perform the actions: S1 – O1, S2 – O2. They are then allowed
to perform the two actions in extinction (without reward) but in the presence
of the two stimuli. In this situation, the animals will perform the two actions
at a low baseline rate. When one of the other stimuli is presented, however,
this leads to an interruption in the animals’ performance and causes them to
immediately shift performance to the action which, in training, earned the goal
or outcome predicted by the stimulus (specific PIT). General PIT occurs when
the stimulus reminds the subject instead of the valence of the goal and simi-
larly may enhance responding nonspecifically. Probably even more relevant
to intrusive thinking is negative or aversive PIT in which the occurrence of a
negative Pavlovian conditioned stimulus (associated, e.g., with painful electric
shock or loud white noise) enhances avoidance or withdrawal behavior or dis-
rupts appetitive behavior, as in conditioned suppression.
These findings suggest that the presentation of the stimulus during the task
might intrude on the animals’ ability to perform their continued action. In this
sense, the stimulus might be regarded as an intrusion, bringing to mind an
outcome. The occurrence of the interruptive stimulus reminds the animal of its
associated action, the performance of which provides a readout of that intru-
sion. PIT is not influenced by extinction of the predictive learning nor by the
devaluation of the instrumental outcome. In these cases, the stimuli continue to
drive the performance of a specific action, demonstrating continued stimulus-
mediated retrieval of the specific, now unwanted, outcome.
Working Memory
While much research relevant to intrusive thoughts, such as those that occur
in PTSD, has focused on long-term memory formation and retrieval (the topic
of the next sections), significantly less work has been done on the role that
working memory mechanisms may play in intrusive thoughts. Yet, by their
very nature, intrusive thoughts can only be intrusive to the degree that they can
gain access to working memory and hijack an individual’s attention. Working
memory is both a capacity-limited and time-limited buffer that allows infor-

mation to be held online in service of current goals or objectives. As with
long-term memory, there are distinct general processes that may be relevant to
intrusive thoughts. Each of these processes is discussed in more detail below
as well as by Banich (this volume).
The first process allows an intrusive thought to gain access to working mem-
ory or to be gated into working memory. Thoughts may be intrusive when they
have priority for access to working memory. Information that enters working
memory may be drawn from the external or the internal environment (e.g., cur-
rent thoughts, long-term semantic or episodic memory, habitual actions), and
may do so via its salience. If the information is drawn from the external world,
it may be salient due to perceptual characteristics; if drawn from either the ex-
ternal world or internal milieu, it may be salient due to conceptual or abstract
characteristics. For example, negative information may be quite salient for in-
dividuals who are depressed, and drug-related information may be quite salient
for those with SUDs. Information may also gain access to working memory
through a more controlled selection process (Feldmann-Wüstefeld and Vogel
2019). Individual differences in the efficiency of such selection processes have
been observed (Vogel et al. 2005), raising the possibility that processes which
allow for selection of information into working memory are altered in indi-
viduals who experience intrusive thoughts.
In the second process, an intrusive thought needs to be given priority within
working memory so that it dominates current thought and actions. Thoughts
may be intrusive when they are stronger and “out compete” other items within
working memory, so being at the focus of attention even when they are not
relevant for the goals or processes at hand. Generally, it is assumed that this
selection process requires cognitive control, and one that can be examined, for
example, using the emotional Stroop paradigm in which individuals are shown
a bivalent stimulus: one feature is task relevant and should guide responding,
the other is not. For instance, an emotional word (e.g., “joy”) is overlaid on a
face with an emotional expression, and individuals must respond based on the
emotional valence of the word (positive, negative). When the word and the
face are incongruent (e.g., the word “joy” on an angry face), reaction time and
error rates are higher than when the items are congruent (e.g., the word “joy”
on a smiling face). This effect is thought to arise because the task-irrelevant
information, the face, is salient—so much so that it interferes with processing
of the task-relevant information (in this case, the word). At least some part of
the interference observed for incongruent as compared to congruent trials is
thought to arise at the level of working memory representations (e.g., Banich
2009). In this paradigm, the salient information is provided by its stimulus
characteristics (i.e., being a word not a face). In other cases, such as the “recent
negatives,” task information can be selected based on its temporal tag. Thus,
individuals are shown a series of four items on a screen and, after a brief de-
lay, a probe prompts them to decide if that item appeared in the immediately
preceding set of four. On certain trials, the probe is one of the four items from
the prior trial, but not the current one. In such cases, individuals show reduc-
tions in performance compared to a less recent probe. This paradigm demon-
strates that time can be a marker for controlling what information is within the
current focus on attention. Future research might explore the degree to which
selection (either on the basis of perceptual, topical, or temporal characteris-
tics) is affected in individuals who experience intrusive thoughts. To the degree
that these control mechanisms are defective, they may not allow for intrusive
thoughts to be moved from the focus of attention, or manipulated or reinte-
grated in such a way as to reduce their saliency (e.g., via reappraisal).
In the third process, information must be removed at some point from work-
ing memory to allow more relevant information to be placed in this limited
buffer. Theoretically it has been argued that removal may occur via three po-
tential mechanisms: (a) passive decay, (b) being replaced by something else,
or (c) being actively removed (Lewis-Peacock et al. 2018). Thoughts may be
intrusive to the degree that they are particularly resistant to such processes.
These processes may be especially important for disorders of intrusive thoughts
which are more abstractly cognitive (i.e., not sensory based) and more repeti-
tive in nature. Moreover, in certain cases it may just be specific categories of
information that are specifically resistant to removal. For example, the degree
of rumination observed in depressed individuals is associated with difficulty
in removing negatively biased information from working memory (Joormann
and Gotlib 2008). This difficulty may occur because of an entrenched overrid-
ing schema for thought that focuses attention to negative material.
Long-Term Memory
Many intrusive thoughts concern contents that have been encoded into long-term
memory. This content includes not only personal autobiographical memories
of past events, but also other contents that clearly rely on declarative memory
but are not autobiographical memories per se (e.g., images of feared future
scenarios or memories of prior thoughts). As noted earlier in this chapter, these
intrusive thoughts are often elicited by retrieval cues in the environment or, in
some cases, cues that are internally generated by the person. While sometimes
positive or neutral, intrusions often have a negative tone. In their extreme form,
intrusive memories involve vivid, multisensory images from highly aversive
events, constituting the hallmark symptom of disorders such as PTSD. When
looking at the role of mnemonic processes underlying intrusive memories and
opportunities to intervene, there are different stages of memory that need to be
considered. A distinction can be made between the encoding of a memory and
its retrieval: within the context of encoding, processes of encoding suppression
can limit the impact of an event; within the context of retrieval, one can focus on
opportunities for memory modification as well as mechanisms for suppression.
Critically, intrusive thoughts involving content encoded into long-term memory
(whether autobiographical, future scenarios, or memories of prior thoughts) can

be conceptualized as involuntary retrieval episodes.
Memory Encoding
Some evidence suggests that the later intrusiveness of an emotional mem-

ory is influenced by processes at the time of encoding and shortly thereafter.
Although emotional intensity contributes to recurring intrusive thoughts for
a variety of reasons, the impact of emotion on both the strength and nature
of the memory that is encoded may be especially important. As reviewed
later in the chapter (see section, What Are the Neural Systems Relevant for
Intrusions and Their Control?), abundant evidence has documented interac-
tions between the amygdala and the hippocampus during emotional events,
putatively contributing to a strong encoding that makes a memory more
prone to later retrieval. Indeed, one view is that intrusive emotional memo-
ries are simply stronger memories, making them more accessible and eas-
ily triggered (Berntsen and Rubin 2008, 2013; Rubin et al. 2016b). In line
with this, emotional memories are typically vivid and are often particularly
resistant to the passive processes that usually lead memories to be forgotten
over time (Hamann 2001); active forgetting processes, however, may behave
differently, as discussed next. Another view is that memories for the specific
episodes of which elements intrude may actually be worse; that is, they are
less contextualized and therefore more fragmented (Brewin 2016; for a chal-
lenge to this view, see Rubin et al. 2016a; Bisby and Burgess 2017). This has
led to the proposal of a “dual representation” of traumatic memories, with
a hippocampus-based system underlying the neutral, declarative aspects of
a memory, and a system involving the amygdala and sensory areas repre-
senting sensory, emotional aspects of a memory (Brewin 2014; Bisby and
Burgess 2017). Although this may be an oversimplification, there is evidence
that the intrusiveness of a memory can be selectively reduced while leaving
the voluntary recall of a memory intact (Holmes et al. 2009, 2010; James et
al. 2015; Gagnepain et al. 2017; Lau-Zhu et al. 2019) and vice versa (Bourne
et al. 2010). This finding is compatible with the possibility that these types of
memories may rely on distinct systems (Visser et al. 2018), or at least distinct
representations (Gagnepain et al. 2017).
Whereas emotional intensity at encoding can amplify a memory’s intru-
siveness, this can often be mitigated by control processes that limit encod-
ing and consolidation. For example, studies using experimental models for
trauma show that engaging in a competing visuospatial task, during or shortly
after experiencing an event, can reduce the frequency of intrusions of that
event over the subsequent week, as evidenced through self-reports in a sub-
ject’s daily diary (Holmes et al. 2004; 2010; Lau-Zhu et al. 2019). In addi-
tion, a sizeable body of work on item-method directed forgetting shows that
when people are instructed to forget the immediately preceding memory item
in a list, the instruction has a large impact on later retention of the memory,
rendering to-be-forgotten items significantly less accessible, even for emo-
tionally unpleasant scenes (Anderson and Hanslmayr 2014). Such directed
forgetting effects arise not only on recall tests, but also on recognition and
implicit memory tasks, indicating that participants can successfully block
the encoding of recent events into long-term memory when motivated to do
so. This encoding suppression mechanism is likely one coping process that
healthy individuals use to limit the impact of upsetting events, reducing their
intrusiveness (Anderson and Hanslmayr 2014). It may also be the key pro-
cess underlying the previously discussed phenomenon of mnemic neglect
(Sedikides et al. 2016).
Memory Retrieval: Modification
Sometimes the encoding of unwanted content occurs despite a person’s efforts

to prevent it. When this happens, people are vulnerable to later involuntary
retrievals when reminders occur in the environment. During these retrievals,
several processes become relevant that could modify the memory and play
a critical role in whether intrusions develop or are mitigated. Understanding
these memory modification processes presents opportunities for intervention,
even for memories that have been successfully consolidated (see Figure 9.1).
Much of the research on memory modification has been conducted within the
tradition of classical conditioning in experimental animals, although memory
modification work also extends to declarative memory.
In research on memory modification, retrieval or recall of a previously consol-
idated memory has been proposed to engage two possible mechanisms: extinc-
tion and reconsolidation (Suzuki et al. 2004; Lee et al. 2006; Clem and Schiller
2016). In extinction, the repeated presentation of a conditioned stimulus leads to
the progressive decrease in, and a resultant reduction in, the behavioral expres-
sion of a memory. In reconsolidation, the presentation of an isolated retrieval is
thought to initiate a molecular cascade that can be bidirectionally modulated to
either strengthen or weaken a memory trace (Dudai 2004; Suzuki et al. 2004;
Lee 2009). For example, blocking de novo protein synthesis or other cellular and
molecular processes critical for memory destabilization and reconsolidation pre-
vents subsequent memory expression (Nader and Hardt 2009; Flavell et al. 2011;
Elsey et al. 2018; Orederu and Schiller 2018).
Studies utilizing a noninvasive behavioral approach combined principles
of reconsolidation and extinction to update a conditioned fear memory.
Specifically, one day after rats acquired associative memories through fear
conditioning, they were presented with an isolated retrieval trial; then, within
the reconsolidation window, they received an extinction session. This para-
digm led to an enduring decrease in conditioned responding, which unlike
standard extinction (i.e., extinction not preceded by an isolated retrieval trial)
did not result in the return of fear as assessed via renewal, reinstatement, and
LTM encoding and

consolidation
Internal bodily and Working

Long-term affective states memory
memory
Reconsolidation External
Extinction
Retrieval environment
Figure 9.1 Interactive interplay in memory formation, maintenance, and modifica-

tion. When initially acquired, memories are fragile and short lived. Some items or
events may only be required for a brief moment and thus may be only temporarily
maintained in working memory. Others may enter the process of consolidation and be
stored in long-term memory (LTM). Upon retrieval, a consolidated LTM may reenter
working memory and be temporarily put to use or trigger mechanisms of extinction or
reconsolidation.
spontaneous recovery, and led to a retardation of fear reacquisition (Monfils

et al. 2009). The findings provide evidence that retrieval followed by extinc-
tion promotes an updating of the original emotional memory rather than the
formation of a competing memory trace (Cahill and Milton 2019). This idea
was extended to fear conditioning in healthy humans (Schiller and Delgado
2010), individuals with phobias that received in vivo exposure (Telch et al.
2017), as well as a number of other forms of learning in rodents and humans
(Xue et al. 2012; Sartor and Aston-Jones 2014; Luo et al. 2015; Björkstrand
et al. 2016; Germeroth et al. 2017; Lee et al. 2017). In a clever twist, re-
searchers have started to employ a procedure involving a distracting task
(the computer game Tetris) following retrieval to update a traumatic memory
(James et al. 2015; Iyadurai et al. 2018; Kessler et al. 2020; see also Holmes
et al., this volume).
Memory retrievals serve as an opportunity to modify memories and in-
deed, other aforementioned protocols could actually be understood through
the lens of reconsolidation updating (i.e., a change in the memory during a
period of destabilization/restabilization). If intrusions are viewed as involun-
tary retrievals, they may provide a window of opportunity to modify an un-
wanted memory (see Figure 14.3 in Holmes et al., this volume), using either
extinction mechanisms with extended exposures (Cassini et al. 2017; Hu et
al. 2018), reconsolidation updating (with extinction, cognitive restructuring,

a distracting task or pharmacological disruption), or suppression mechanisms
(described below). It may be important to consider that failure to act in the
face of pathological intrusions could lead to their reinforcing a memory trace.
Alternatively, the unique nature of intrusions may represent a form of mem-
ory that is resistant to modification. For instance, it may not destabilize upon
retrieval or fail to promote extinction or reconsolidation processes. In this
case, more targeted behavioral or pharmacological interventions (or a combi-
nation of the two) may help.
Memory Retrieval: Suppression
As discussed, intrusions can be conceptualized as involuntary retrievals,

whether they are autobiographical memories, images of feared future events,
or persistent thoughts. Efforts to control intrusions therefore involve control-
ling retrieval. Here, we discuss two processes that contribute to controlling
such thoughts: retrieval suppression and retrieval-induced forgetting.
Retrieval suppression refers to the act of trying to stop an ongoing retrieval
process, usually triggered by a reminder (Anderson and Green 2001). For in-
stance, a loved one’s face may elicit an intrusion of a recent argument that
caused upset or a dog barking may trigger memories of the night a person got
news of a friend’s death. Other retrievals may be of a feared future: a medical
bill may remind someone of an upcoming medical test. Because people do
not enjoy the aversive emotional states that such reminders can create, they
often exclude the offending content from awareness to regain their focus and
composure. The aim is not merely to stop the intrusive thought at the moment,
but to diminish its recurrence. Before describing retrieval suppression, we first
discuss differing views on the value of suppression and whether it is desirable
to encourage.
The Utility of Suppression. The clinical psychology literature contains con-

flicting observations on the value of thought suppression. Many clinicians,
particularly those concerned with trauma or anxiety, maintain that suppres-
sion is an intrinsically unhealthy response to intrusive thoughts or memories.
Indeed, it is common to view suppression as not only unhelpful, but as a sig-
nificant risk factor for developing psychiatric disorders. Some studies have
found that questionnaires putatively measuring thought suppression (Wegner
and Zanakos 1994) are related to worse symptom severity in PTSD and other
disorders characterized by intrusions. Given these facts, one can readily under-
stand why clinicians would conclude that suppression is maladaptive. As a re-
sult, many therapists and therapies (e.g., acceptance and commitment therapy)
discourage suppression and encourage people to accept and interact with their
intrusive thoughts.
Several observations, however, are incompatible with the view that sup-
pression is maladaptive. First, suppressing unwelcome thoughts is a wide-
spread behavior that serves many different functional purposes, as outlined
above. Simply concentrating on an important task, by its nature, requires that
an individual excludes distracting content. Second, achieving emotional bal-
ance after an upsetting event involves “getting over” unwelcome thoughts and
feelings, a process that requires active regulation of thoughts. People who have
difficulty with upsetting occurrences (e.g., getting over arguments or perceived
slights) are viewed as coping less well than people who quickly recover. Third,
after a trauma, most people initially experience intrusive thoughts and memo-
ries that usually diminish over time. People who cannot reduce their intru-
sions are classified as having mental health concerns; those who are capable
of reducing intrusive thoughts effectively are considered resilient, a clearly
positive attribute. Fourth, people with anxious thoughts about a feared event
that they cannot discard are considered as less healthy, compared to those who
are able to set aside fears and cope well. Finally, the proposal that trying to
stop unwanted thoughts is intrinsically ineffective is oddly discordant with the
massive literature on inhibitory control and, more broadly, attentional control,
in which controlling cognition and behavior is not only desirable but an essen-
tial capacity of intelligence. Indeed, the same clinicians that classify suppres-
sion as maladaptive often maintain beliefs that are contradictory to that stance
upon closer inspection. For example, the belief that attention bias modification
(i.e., training people to ignore unpleasant interpretations) is a desirable therapy
amounts to a belief that it is desirable to become skilled at ignoring unpleasant
contents, exactly what people do through suppression. The belief that people
with intrusive thoughts suffer from cognitive control deficits that make them
vulnerable to such thoughts assumes that resilient people can set aside upset-
ting thoughts, which must therefore be beneficial.
What do we make of these contradictions? What accounts for the view that
suppression is maladaptive, given its ubiquity and utility? One reason derives
from work on thought suppression using the white bear paradigm (Wegner
et al. 1987). In this paradigm, participants are instructed not to think about a
particular thought (e.g., a white bear) over a five-minute period. During that
time, if they happen to think about the thought, they were asked to report this.
Afterward, they are given a five-minute free expression period. Wegner found
that participants asked to suppress thoughts of a white bear ironically experi-
enced more white bear thoughts than participants who were not asked to sup-
press. This enhancement of the unwanted thought after suppression was termed
the rebound effect (Wenzlaff and Wegner 2000). Across nonclinical samples,
meta-analyses show that people exhibit a small to medium rebound effect com-
pared to control instructions (Abramowitz et al. 2001). Wegner and Zanakos
(1994) also introduced the White Bear Thought Suppression Inventory, a scale
intended to identify people prone to thought suppression. This instrument be-
came widely used in clinical research and was found to correlate with both
rebound effects and clinical symptomatology. Coupled with this question-

naire, the white bear paradigm provided a compelling narrative that explained
why psychopathologies were associated with intrusive thinking: unpleasant
thoughts led to suppression, which led, in turn, to a paradoxical rebound effect,
yielding a vicious cycle that worsened psychological symptoms (Wegner and
Zanakos 1994).
There are significant difficulties with treating the white bear task as a model
for controlling intrusive thoughts. First, the task requires the integration of the
very thought to be suppressed with the task goal; that is, to know what task one
is doing and to monitor whether one is achieving it, one reactivates the thought
by asking: “Am I thinking about white bears?” The task is self-defeating be-
cause it requires one to violate the task’s goals to check whether the goal is
being accomplished (Anderson and Huddleston 2012; Engen and Anderson
2018). Although the situation modeled by this task may model a slice of clini-
cal reality, work using this paradigm has been overgeneralized to all instances
of suppression (as will be seen shortly). Second, the White Bear Suppression
Inventory—thought to quantify thought suppression frequency—has now been
shown to have two factors opposed to one another: one measures thought sup-
pression frequency while the other, the experience of intrusiveness (Blumberg
2000; Höping and de Jong-Meyer 2003; Rassin 2003). Research indicates that
only the latter correlates with clinical symptoms, which is not surprising. Third,
meta-analytic treatments demonstrate that suppression, as measured with the
white bear task (and its rebound effects), shows few reliable differences across
control and psychiatric populations, raising questions about the validity of the
relationship of the process being measured and disordered control over intru-
sions in everyday life (Magee et al. 2012). Finally, the clinical observations that
led clinicians to be attracted to the claim that thought suppression is counterpro-
ductive—reduced suppression in patients—can be explained more simply by
positing that patients have difficulties applying inhibitory control rather than by
suppression being intrinsically maladaptive. In line with the former view, sup-
pression is an otherwise healthy coping response which, when disordered, poses
a risk factor in developing intrusive symptoms—a very different interpretation
with different clinical implications. For example, if suppression is not intrinsi-
cally maladaptive, but simply not functioning properly in some populations, it
is reasonable to expect that interventions can be developed to improve it. This
possibility would be ruled out by the view that suppression is maladaptive.
Another reason why some clinicians have been opposed to suppression in-
volves confusions in terminology. Researchers have not distinguished retrieval
suppression from other constructs that may well be maladaptive and that might
appear equivalent to retrieval suppression. For example, researchers have
falsely equated retrieval suppression with the emotion regulation construct
of expressive suppression and other seemingly related terms, such as avoid-
ance, cognitive avoidance, and distraction. Careful analysis of the situations
captured by these terms reveals that retrieval suppression is cognitively dis-

tinct from these other concepts:
• White bear suppression (Wegner et al. 1987) refers to the explicit inten-
tion to not think about a specific unwanted thought for a specific period
of time (e.g., 5 min). The unwanted content is incorporated as part of
the goal and integrated with it, making it, by definition, not possible to
avoid periodic attention to the thought. In contrast, retrieval suppression
involves the attempt to stop or cancel the retrieval of content associated
with a reminder while sustaining attention to the reminder. A person’s
goal does not specify the content to be inhibited but rather concerns the
desire to shut down “whatever is associated with the reminder.”
• Expressive suppression (Gross and John 2003) refers to the strategy of
inhibiting behaviors associated with emotional states, such as adopt-
ing a poker face to hide emotions. Expressive suppression involves
motor control, not control over thoughts or feelings. Retrieval sup-
pression, on the other hand, involves suppressing mnemonic processes
and content.
• (Cognitive) avoidance (Ehlers and Steil 1995; Williams and Moulds
2007) refers to entirely avoiding reminders that could otherwise trig-
ger unwanted thoughts or emotions. It fundamentally involves not con-
fronting reminders and not adapting one’s internal response to them
so that unwanted contents are preserved without alteration and may
continue to intrude at the slightest provocation. It is widely considered
a maladaptive coping process, a “not dealing with the problem.” In
contrast, retrieval suppression requires a person to confront reminders
directly, attend to them, and adjust their internal retrieval and affective,
conditioned responses to them, features shared with cognitive behav-
ioral therapy and exposure therapy in particular.
• Distraction refers to the removal of attention from emotions or thoughts
and refocusing it onto other innocuous stimuli in the world or to un-
related thoughts. It is worth distinguishing general distraction (where
attention shifts to stimuli, topics, or activities unrelated to the intrud-
ing thought) from specific distraction, which involves interacting with
reminders. If general distraction leads one to remove attention from
reminders to intrusive thoughts or emotions, it functions in a manner
similar to avoidance and fails to engage suppressive processes that
might adjust one’s internal mnemonic and emotional responses. If dis-
traction takes the form of generating alternative associations, thoughts,
or memories in response to a reminder, this form of specific distraction
would constitute thought substitution, which retrains the response to
the cue to elicit different content, potentially leading to altered patterns
of thought in the future. Sometimes, because retrieval suppression in-
volves attending to reminders and not the memory, or instead, attention
to thought substitutes, researchers falsely equate retrieval suppression

with general distraction, from which it clearly differs.
Because clinical research has often not attended to these distinctions, it is easy
to see how, with the paradoxical rebound effects reported in the white bear
paradigm, clinicians would have taken a dim view of suppression. In recent
years, retrieval suppression has been studied intensively. As we look at next,
this work integrates models of the control over intrusive thinking within the
bedrock of cognitive neuroscience research on inhibitory control.
Retrieval Suppression: Methods of Study and Basic Findings. Work on re-

trieval stopping focuses on a situation that is similar to motor response inhibi-
tion. Participants are confronted with a cue that reminds them of a memory and
are asked to focus on that cue for trials of several seconds, while preventing
the memory from entering awareness. This situation parallels motor response
inhibition tasks, such as the Go/No-Go and Stop-Signal tasks, in which par-
ticipants are presented with stimuli designed to elicit a motor response, yet are
required to stop the action. To study retrieval suppression, Anderson and Green
(2001) introduced the Think/No-Think paradigm. Participants are trained on
cue-target pairs until they can recall the target items when given the cues.
These pairs can be unrelated words (e.g., ordeal roach), picture pairs, or even
cues related to autobiographical memories. Participants then enter the Think/
No-Think task. On each trial, participants receive a reminder from one of the
pairs, colored either in green or red. Green reminders signal the participant
to retrieve the associated item and keep it in awareness for the trial’s dura-
tion; red reminders signal that they should attend to the stimulus but prevent
the associated item from entering awareness at all. Usually, a given item is
only suppressed or retrieved (not both), and will be repeated many times (usu-
ally, 8–16 times). A final memory test follows, giving participants all studied
cues to determine the impact of retrieval suppression on the suppressed con-
tent. Performance on the final memory test is computed not only for the Think
items, which were repeatedly retrieved, and No-Think items, which were sup-
pressed, but also for baseline pairs studied in the training phase but which did
not appear in the Think/No-Think phase.
A key finding that emerged from this work is that if people consistently sup-
press the retrieval process, memory for the suppressed No-Think items declines
compared to Think items as well as to baseline items receiving the same initial
training as No-Think items. Because suppressing retrieval impairs retention
of the suppressed content, this phenomenon is known as suppression-induced
forgetting (for a review, see Anderson and Hanslmayr 2014). Suppression-
induced forgetting increases with the number of times that people suppress the
unwanted content and resists monetary incentives on the final test for success-
ful recall (Anderson and Green 2001; Hulbert and Anderson 2018). Forgetting
generalizes to novel test cues for the suppressed items, exhibiting a property
known as cue independence: suppression-induced forgetting occurs not only

for simple word pairs, but also for objects or complex scenes, irrespective of
whether they are neutral or negatively valenced (Depue et al. 2007; Anderson
and Hanslmayr 2014). Suppression-induced forgetting has even been observed
with autobiographical memories, although the effect emerges more in memory
for details (Noreen and MacLeod 2013). Importantly, a growing body of work
shows that suppression-induced forgetting does not merely influence explicit
memory, it also affects implicit memory regardless of whether perceptually
driven (Kim and Yi 2013; Gagnepain et al. 2014) or conceptually driven
tasks are used (Hertel et al. 2012, 2018; Taubenfeld et al. 2019). Of clinical
relevance, retrieval suppression also reduces affective responses to the sup-
pressed content when participants view it later, whether measured with ratings
(Gagnepain et al. 2017) or psychophysiology (Legrand et al. 2019; Harrington
et al. 2020). These findings suggest that retrieval suppression, for healthy par-
ticipants at least, regulates the negative affect associated with intrusive memo-
ries (Gagnepain et al. 2017; Engen and Anderson 2018).
Although most research on retrieval suppression focuses on final test per-
formance, recent studies have increasingly focused on online measurements
of intrusive memories during the Think/No-Think phase (Levy and Anderson
2012; Benoit et al. 2015; Hellerstedt et al. 2016; van Schie and Anderson 2017;
Legrand et al. 2019; Harrington et al. 2020). In these studies, the Think/No-
Think task is unchanged except that after every trial, participants provide a
phenomenological report of their experience. Specifically, they judge whether
the associated memory entered awareness. On Think trials, awareness of the
memory is the intended outcome, and people usually report awareness on
nearly all trials (averages of 90–100% of trials). On No-Think trials, however,
because participants are instructed to prevent awareness of the memory, any
awareness that does occur constitutes clear evidence for an intrusive thought.
Importantly, when intrusions occur in this task context (e.g., on 30% of No-
Think trials), the inference of involuntariness is particularly clear. Indeed, one
of the hallmark features of an involuntary, automatic behavior is its tendency
to occur, despite efforts to stop it. Intrusions during No-Think trials therefore
provide a robust and theoretically clean measure of a vulnerability to involun-
tary retrieval. Other paradigms hoping to measure intrusions (e.g., diary stud-
ies) cannot make a clear attribution of involuntariness (for discussion, see van
Schie and Anderson 2017).
Contrary to what might be inferred from the white bear paradigm, people
can suppress the retrieval process and reduce intrusive thoughts elicited by
reminders. For example, across repeated suppressions one observes a dramatic
downregulation of intrusions between the initial trials (e.g., 60% intrusions)
and final trials (e.g., 25% by the tenth trial). Moreover, the slope of the down-
regulation in intrusions is often correlated to suppression-induced forgetting
(Levy and Anderson 2012; Hellerstedt et al. 2016), and suppression effects per-
sist to produce memory deficits on both direct and indirect tests, contrary to the
notion of rebound. Work on retrieval suppression aligns the control of intrusive

thoughts with work on inhibitory control in general, especially work on action
stopping. Indeed, given that nobody questions that actions can be stopped and
that a giant literature has developed which documents action-stopping mecha-
nisms in humans and other animals, it would be more surprising to find that in-
ternal actions, such as retrieval, were not also subject to control (Anderson and
Green 2001). This conclusion fits the observation that psychiatric populations
with intrusive thinking also have deficits in cognitive control (Catarino et al.
2015; Waldhauser et al. 2018). Suppression-induced forgetting has also been
related to self-reports of rumination (Fawcett et al. 2015), perceived success
at thought control in real-life settings (Küpper et al. 2014), and self-reports of
intrusive memories collected in the week following exposure to a trauma film
(Streb et al. 2016). These findings support the view that suppression is a help-
ful coping process that is compromised in clinical populations, rather than an
intrinsically maladaptive approach to intrusive thoughts.
Retrieval-Induced Forgetting: Methods and Basic Findings. Thus far we

have focused on the role of inhibitory control in intentional retrieval stopping.
Inhibitory control, however, can also be engaged in another context to forget
distracting memories: selective memory retrieval. During selective memory
retrieval, a person seeks to retrieve a particular event, idea, or fact. Reminders,
however, are often associated to many traces, creating interference known
as retrieval competition. Retrieval interference is ubiquitous, whether one is
retrieving an episodic memory, a fact, a general idea, or simply an object’s
location (Anderson and Neely 1996). Isolating the desired trace despite this
interference engages prefrontally mediated inhibitory control mechanisms to
suppress competing memories. This suppression induces aftereffects on the
inhibited traces causing forgetting. Thus, ironically, the very act of remember-
ing something can cause forgetting of competing memories; this is known as
retrieval-induced forgetting (Anderson et al. 1994; Anderson and Spellman
1995; Levy and Anderson 2002; Anderson 2003; Bäuml et al. 2010; Storm and
Levy 2012). Although retrieval-induced forgetting is not intentional, this pro-
cess can be exploited in a motivated manner to deliberately forget unwelcome
memories (Benoit and Anderson 2012). The procedure by which this phenom-
enon is studied will be described below.
Retrieval-induced forgetting becomes particularly relevant to controlling
intrusive thoughts when it is deployed in a motivated way; that is, as a key
strategy for deflecting unwelcome thoughts, refocusing on benign knowledge.
For example, after an argument with a friend or partner, one might (upon see-
ing them again) try to recall more pleasant thoughts about them (Anderson
2001). Optimists routinely seek the positive in any event, even if unpleasant
things happened, and in so doing, may reshape their memories to fit a more
pleasant existence. Indeed, the widely demonstrated positivity bias in auto-
biographical memory retrieval (the disproportionate accessibility of positive
over negative autobiographical memories) significantly correlates with the

amount of retrieval-induced forgetting that people show on laboratory tests
with neutral materials (Storm and Jobe 2012). Thus, the capacity to exhibit in-
hibitory control over memory predicts a sunnier view of one’s past than might
be expected. Critically, retrieval-induced forgetting may be a mechanism that
explains why reappraisal is effective as an emotion regulation strategy, given
that reappraising an unpleasant memory involves selectively retrieving the
contents of an event or thought (Engen and Anderson 2018). If conventional
therapies for addressing intrusive thinking rely on reframing, rescripting, or
reappraisal, retrieval-induced forgetting may be central to reducing the acces-
sibility of negative interpretations and intrusions.
Selective retrieval’s role in motivated forgetting has been studied with
the retrieval-practice paradigm as well as with a variant of the Think/No-
Think paradigm, known as thought substitution (Hertel and Calcaterra 2005;
Bergstrom et al. 2009; Joormann et al. 2009; Benoit and Anderson 2012). In
thought substitution, participants are instructed to stop retrieval of the un-
wanted memory by retrieving diversionary content, either self-generated or
supplied, related to the cue (Hertel and Calcaterra 2005; Bergstrom et al. 2009;
Benoit and Anderson 2012). Retrieval of such thought substitutes significantly
impairs retention of the intrusive thoughts, although the neural mechanism un-
derlying this process is dissociable from that involved in retrieval suppression
(Benoit and Anderson 2012). Thought substitution as a strategy for controlling
intrusive memories has been studied extensively using the Think/No-Think
task, especially in dysphoria and depression (Joormann et al. 2009; Stramaccia
et al., unpublished). Given people’s natural pre-disposition to replace intrusive
contents with distracting thoughts, how thought substitution affects memory is
important to study.
Most research on intrusive thoughts has focused on long-term memory forma-

tion, retrieval, and their control. The substrate for intrusive thoughts is laid
down when memories are encoded. Many intrusive thoughts are related to past
experiences, such as those associated with substance use or specific events
(e.g., feeling rejected, being traumatized). Other intrusive thoughts may not
refer to specific episodic memories (e.g., images of feared future events or
ruminations) but are nonetheless stored in memory and governed by its mecha-
nisms. If a memory is encoded in a particular way, it may be more accessible
to reactivation at a later time. Highly salient memories may, in particular, be
easily triggered by external and internal cues, many of which do not reach
conscious awareness. Intrusions, therefore, appear to pop spontaneously into a
person’s mind. Such potent memories may be related to specific types of emo-
tional information (e.g., negative information as in depression and anxiety) or
to a particular topic (e.g., danger and threat in PTSD and anxiety; substance-
related information in substance use disorders).
Substantial research has focused on cognitive control processes, including
the role of inhibitory control mechanisms in intrusive thinking. Memory con-
trol mechanisms related to action control are known to suppress retrieval of
specific, episodic long-term memories and can work more globally to block all
episodic memory retrieval in response to a cue. Alternatively, reminder cues
can be used to selectively retrieve alternative distracting content, suppress-
ing the intruding thought via the mechanisms of retrieval-induced forgetting.
Moreover, retrieving information may trigger destabilization and/or restabi-
lization of a thought. As such, retrieval provides an opportunity not only to
increase the strength of a memory, but also to reduce it. In addition, working
memory processes are critically linked to the encoding and retrieval of long-
term memory and are important to guiding the selection of information for
consolidation into, and retrieval from, long-term memory (see Figure 9.1). It
seems likely that for an intrusive thought to be experienced, it would enter
working memory, making this system central to having an experienced (con-
scious) mental event.
Summary of Commonly Used Methods and

Desiderata for Investigating Intrusive Thinking
Methods and Paradigms
The previous section described examples of research into the mechanisms un-
derlying intrusive thinking and their control using a variety of methods. Here
we present an overview of the methods and paradigms that are most commonly
used to study intrusive thinking in humans, nonhuman animals, or both, each
with its own strengths and weaknesses.
• Experience/thought sampling is a form of self-report that is of-
ten used to assess mind wandering in the present moment (Larson
and Csikszentmihalyi 1983; Bolger et al. 2003; McVay et al. 2009;
Schooler et al. 2011; Baird et al. 2014; Fraley and Hudson 2014). One
form has people pressing buttons when they catch themselves mind
wandering or experiencing other intrusions. This can be done in the lab
or in real life (e.g., using a diary or smartphone). Alternatively, one can
also probe people at irregular intervals to report whether they were on
task or not. A comparison of probed versus unprobed self-report shows
that having an experience and knowing that you are having an experi-
ence are different things (Schooler et al. 2011; Baird et al. 2014). The
first requires metacognition and this is imperfect, and possibly worse
in clinical populations.
• In the emotional Stroop task, individuals view items with two

overlapping dimensions, such as a word superimposed on a face
(Williams et al. 1996). The participant bases their response on a
task-relevant dimension (e.g., valence of a word) while ignoring the
task-irrelevant dimension (e.g., valence of a face). Proneness to in-
trusions is defined as the degree to which reaction times or error
rates are increased on incongruent trials (e.g., the word “suicide” on
a happy face) as compared to congruent trials (e.g., the word “joy”
on a happy face).
• The trauma film paradigm is used to model intrusive memories of dis-
tressing events (Horowitz 1969; James et al. 2016a). It uses film stimuli
in the laboratory, which contain traumatic content that can bring about
intrusive memories subsequently in daily life. These memories are typ-
ically recorded in a diary, allowing for a frequency count of intrusive
memories, or via button presses during a provocation task (both are
forms of experience sampling).
• Pavlovian conditioning is the classic model by which simple as-
sociative learning and memory are studied, and it is well suited for
research across species (Pavlov 1927; Rescorla and Holland 1982;
LeDoux 2003; Nader and Hardt 2009). In this paradigm, an initially
neutral stimulus (conditioned stimulus, CS+), such as a triangle, is
repeatedly paired with an intrinsically aversive stimulus (uncondi-
tioned stimulus, UCS), such as an electric shock, while another con-
ditioned stimulus (CS–), such as a circle, is never paired with the
UCS. With sufficient CS+/UCS pairings, the CS+ acquires the same
aversive qualities as the UCS and will elicit a conditioned defensive
response that can be measured, for example, by the amount of freez-
ing or avoidance behavior in nonhuman animals, or skin conductance,
acoustic startle response, heart rate, pupil dilation, action tendencies,
UCS expectancies, and subjective distress in humans. After repeated
presentations of the CS+ without the UCS, the defensive response
usually diminishes, a process referred to as extinction learning.
Intrusions may often be conceptualized as conditioned responses to
external or internal reminders.
• Stop-Signal and Go/No-Go are tasks that require inhibition of already
initiated responses. Error rates in response to Stop-Signal or No-Go tri-
als may reflect the outcome of diminished control of actions in a similar
way as intrusions reflect a diminished control of thoughts.
• In the Think/No-Think paradigm, participants learn a set of cue-target
pairs, either word pairs, picture pairs, or autobiographical memory cues
(Anderson and Green 2001). Next, trials appear, each presenting a re-
minder cue and a task cue: a green box around the reminder cue signals
that the associated memory item needs to be brought to mind whereas
a red box signals that retrieval of the associated memory item needs to
be suppressed. This method allows one to measure, during the suppres-

sion task itself, whether participants experienced intrusions (i.e., intru-
sion reports during the No-Think trials) in response to the reminders,
despite their best efforts to stop retrieval.
• Retrieval-induced forgetting is studied using the retrieval-practice par-
adigm in which participants learn to associate a cue with multiple as-
sociates; participants are then set the task of retrieving a subset of these
items, but not the rest (Anderson et al. 2000; Bekinschtein et al. 2018).
Rather than providing explicit instructions to forget, this paradigm cap-
italizes on a natural process to select relevant information. It leads to
forgetting of nonselected, competing memory items through a process
of inhibition. Retrieval-induced forgetting is related to a variant of the
Think/No-Think paradigm that uses thought substitution rather than
retrieval suppression, and when deployed in a motivated way, may be
a model for how individuals control and diminish intrusive thoughts.
Retrieval-induced forgetting is now also being studied in rodents,
capitalizing on rats’ intrinsic curiosity (spontaneous object recogni-
tion paradigm). After encoding multiple objects in an environment,
selective retrieval of one object leads to forgetting of other objects in a
manner directly analogous to that seen in humans and that depends on
the prefrontal cortex. An advantage that this paradigm enjoys over re-
trieval suppression is that it allows us to understand memory inhibition
mechanisms, not only at the level of neural systems (through imaging
in humans) but also through foundational neurobiological work in non-
humans using electrophysiological, lesion, optogenetic, and molecular
biological methods.
• With Pavlovian-instrumental transfer, a subject undergoes Pavlovian
threat conditioning and is also trained on an instrumental reinforcement
task (reward or avoidance) in a different context (Campese et al. 2013;
Cartoni et al. 2016; Watson et al. 2018). In the test phase, during ex-
tinction (or baseline performance) of the instrumental task, the CS+ is
presented and its possible rate-altering effects measured as evidence of
negative Pavlovian-instrumental transfer. There are several variations
of this basic procedure, which can also be appetitive in nature and have
general and specific aspects of transfer in relation to the overall goal.
The degree to which the presentation of the CS+ interrupts an animal’s
(human or nonhuman) goal-directed behavior is taken as a proxy for
the experience of an intrusion.
These paradigms should continue to be developed. In addition, we wish to
stress that promising new approaches are emerging from computational neuro-
science, as we illustrate next.
Computational Approaches to Intrusive Thinking
We turn now to a few considerations from a computational perspective. A com-

mon feature of intrusive thoughts is that they occur at the “wrong time.” They
are remembered or imagined contents that are not appropriate for the situa-
tion or the particular problem our minds are focused on at that time. But what
constitutes the “wrong” time? How do we, our minds, or our brains choose
the “right” thoughts or the “right” time for particular thoughts? Are there any
general principles that guide the selection of thoughts, and what might they
be? Computational modeling of decision making provides a framework for
this problem and suggests a number of potential answers that might apply to
different mental illnesses (Huys et al. 2016; Redish and Gordon 2016). At their
core, these accounts rely on the notion that thoughts are expensive and that
organisms have much to gain from deploying them well.
Our starting point is the acknowledgment that most decision problems
faced by humans on a daily basis are so complex that they radically outstrip
our computational abilities. The mismatch between computational demands
and resources requires us to invest our computational resources with wisdom.
Unfortunately, exerting any such wisdom further complicates the problem:
the optimal deployment of limited computational resources is itself a deci-
sion problem that has the same form as the original decision problem, but
which is vastly more demanding (Russell and Wefald 1991). This problem
is displayed in Figure 9.2a. Hence, we are faced with the double challenge
consisting of hard problems and the even harder problem of apportioning our
cognitive resources.
There is, however, a silver lining for the study of thoughts. As the problem
of optimally apportioning computational resources is essentially of the same
nature as the original decision-making problem, the same formalisms used to
study choice can potentially be brought to bear on the problem of thought se-
lection (Lieder et al. 2018a, b). Errors in these processes might, in turn, link the
emergence of intrusive thoughts to well-defined optimal approaches to solving
decision-making problems under resource constraints (Huys and Renz 2017).
Below, we review several computational accounts of thought choice and dis-
cuss how intrusive thoughts could arise.
Affective Biases
A bias toward thinking about negative events can arise through utility-
weighted sampling, a process whereby the samples (in this case, what we
choose to think about) are not just proportional to the probability of the event
but also to its importance. Humans routinely overestimate the probability of
extreme negative events (Lichtenstein et al. 1978), and such biases have long
been viewed as a signature of irrationality. However, they may alternatively
reflect a rational use of limited resources. When many different outcomes are
(a) (b)
(c) (d)
+140
2 1
Action 1
–7
–2
Action 2 0 0
+2
0
Action 3
p(value)
3 6
4 5
value
Figure 9.2 Computational approaches to thought guidance: (a) The meta-reasoning
problem of optimally apportioning limited resources can be formalized as a decision-
making problem over decision-making problems. While the standard decision-making
problem is the tree at the top without thick lines, the meta-reasoning problem is a tree
of such trees, where each branch corresponds to choosing to evaluate one branch (thick
lines) in the original decision problem. Intrusions could relate to a tendency to only
sample one set of options (arrows). (b) Prioritized replay: In this state-space, if the
reward (Euro coin) is received in the bottom right corner, then standard model-free
learning updates only the state immediately adjacent to it. Prioritized replay allows
memories to be reused multiple times to update more distant states; here, propagating
the information about the reward all the way back to the starting state. If the need or
gain functions determining what memories to replay are altered, then this could result
in repetitive replay of the same memory (arrows) as well as a failure to update dis-
tance states (i.e., to integrate the memory with other memories). Adapted after Mattar
and Daw (2018). (c) Guidance of thoughts by habits: Consider the situation where a
valuation system provides distributions for likely values of actions. Here, action 1 is
clearly inferior to the other two, and it appears that action 2 is the best. If the agent
were to invest computational effort into refining the estimates of the values of the ac-
tions, it would be best to examine action 3, as it may be even better than action 2. In
this manner, a goal-directed or model-based system could elaborate on approximations
provided by a simpler valuation system. This would also mean that the goal-directed
thought choice could be (mis)guided by habits. (d) Stopping aversive thoughts: In this
task, participants are extensively trained to learn to navigate a maze where each transi-
tion incurs some gains or losses. When given the opportunity to plan freely, they stop
internal simulations when they encounter one of the salient losses in red. Adapted after
Huys et al. (2015).
possible, it becomes difficult to consider them all, and humans are thought to
simulate instead a few outcomes in their mind and average the results of these
few simulations. If these simulations are too few in number, they are likely
to miss very important outcomes. Simulating in proportion not only to the
probability, but also in proportion to the utility, allows the resulting estimates
to be more robust. That is, the apparent overestimation of extreme events may
help mitigate cognitive limitations and explains a number of apparent irratio-
nalities (Lieder et al. 2018b). Applied to the setting of intrusive thoughts, it
provides one argument for why salient negative events should be simulated
even when they are very unlikely. In fact, the more aversive the event, the
higher the likelihood of simulation, suggesting one path by which the per-
ceived negative valence of a simulated event might increase the frequency
with which it is thought about.
Trauma Replay
A computational process that may be related to intrusive thoughts in PTSD

is that of prioritized replay. Briefly, learning from experience is often slow. A
prominent way of learning from experience through iterative updates of ex-
pectations with prediction errors is particularly slow. Indeed, this is one reason
why it has been considered to be a computational account of habitual learning
(Daw et al. 2005). Figure 9.2b shows that this is because experience at any one
state or stimulus, s only leads to learning at that particular state or stimulus
and is only propagated to adjacent states upon the next transition from those
adjacent states back into state s. Such experience influences “knowledge” only
very locally and does not generalize. One solution to this is to store episode-
like chunks of autobiographical memory and replay them multiple times so as
to spread the effect of any one experience to other states. This can substantially
accelerate learning (Schaul et al. 2016). However, replaying memories is also
costly, and hence the key here is to again deploy the resources (in this case,
which memory to replay) efficiently. Computational models of this process
identify two terms that determine which memory to replay: a need term and a
gain term (Mattar and Daw 2018). The need term captures how likely state s
is to be encountered in the future (Russek et al. 2017). Clearly, using compu-
tational resources to learn about states that will never be visited is not useful.
The gain term captures how much the memory is likely to change behavior.
As this gives a normative account of when to optimally replay particular
memories, it should reflect the tendency to replay both intrusive and non-intru-
sive memories experienced after laboratory induction procedures. It also pro-
vides an interesting window on features of intrusive memories in PTSD. The
gain term depends on the implied change in behavior and, as such, a memory
should only be replayed if it implies a change in behavior. This suggests an
important modulatory role of generalization processes seen across anxiety dis-
orders (Laufer et al. 2016). For instance, replay tendency after traumatic abuse
by a trusted person should be influenced by the perceived importance of this
event for other relationships. The more relevant it is judged for other relation-
ships, the higher the replay tendency should be. More generally, it captures
the notion that excessive negative appraisals of the sequelae of a trauma might
relate to the emergence of PTSD symptoms (Ehlers and Clark 2000).
Obsessions
In the context of OCD, intrusive thoughts exhibit a different quality to those

experienced in PTSD (see Monfils and Buss, this volume). Extensive evidence
suggests that OCD involves a shift from goal-directed toward habitual behav-
ior, with most evidence pointing to an impairment in goal-directed or model-
free control rather than an explicit change to habitual or model-based processes
(Robbins et al. 2012; Gillan et al. 2013, 2015; Voon et al. 2015). How could
an impairment in model-based control give rise to the emergence of intrusive
repetitive thoughts as seen in OCD? One avenue arises from the notion of
value of information (Keramati et al. 2011). Consider the situation in Figure
9.2c, where one option is clearly good and a second one clearly bad. The third
option appears slightly worse than the best one, but could be better. The opti-
mal investment of limited cognitive resources in this case would be to examine
this one option, again because this investment of cognitive resources has the
potential of altering behavior. This could underlie intrusive thoughts in OCD
if there was a drive provided by uncertain habitual or model-free evaluations,
such as through distributional reinforcement learning (Dabney et al. 2017),
coupled with impairments in goal-directed evaluations which fail to result in
improved predictions.
Rumination
Frequently seen in depressive disorders, ruminations are usually described as

a tendency to think repetitively about the causes of distress without engag-
ing in active problem solving (Nolen-Hoeksema et al. 1993, 2008; Treynor et
al. 2003). Viewed in the context of intrusive thinking, they appear to involve
a prominent failure to inhibit or discontinue aversive sequences of thoughts.
This raises a theoretical point not yet addressed above—that of thought inhibi-
tion. Thus far we have focused on which thoughts, evaluations, or memories
should be chosen and have not yet addressed the monitoring of a thought that
appears not to be fruitful nor the question of when it should be terminated.
This question has been examined in some detail using computational
accounts of the task in Figure 9.2d. Here, individuals are trained to navi-
gate a maze where each transition yields rewards or losses. They are then
dropped into one state randomly and asked to search a route of a given
length through the maze that maximizes their total earnings. The problem
they face is a binary decision tree, as in Figure 9.2a. Individuals are much
more likely to identify optimal routes that do not transition through salient
losses than those which do, independently of the size of the large loss (Huys
et al. 2012, 2015). Computational models of the choices in this task suggest
that participants internally simulate potential routes through the maze and
terminate simulation when they encounter a salient loss. Functional mag-
netic resonance imaging suggests that this inhibition recruits the subgenual
anterior cingulate cortex (Lally et al. 2017), a region known to be impor-
tant in depression (Drevets et al. 1997), its treatment (Mayberg 2009), and
rumination (Hamilton et al. 2015). Indeed, individuals who score high on
self-reported rumination show reduced pruning; that is, a reduced tendency
to terminate thoughts when encountering large losses during their internal
simulations (Q. Huys, pers. comm.), suggesting that rumination might di-
rectly relate to an inability to inhibit aversive thoughts, possibly via impair-
ments involving the subgenual anterior cingulate.
Desiderata for More Sophisticated Behavioral Paradigms to

Measure Intrusive Thinking and Its Control
Studying intrusive thinking is challenging for several reasons. Intrusive

thoughts are typically spontaneous, making it hard to predict when they occur,
and therefore, when to measure them. Although paradigms exist that can reli-
ably induce intrusive thinking under controlled circumstances in the laboratory
(as detailed above), additionally allowing for the investigation of its neural
correlates, these paradigms may not capture all circumstances under which
intrusions occur in real environments. In addition, while it is possible to infer
the occurrence of intrusion-like events from nonverbal behavior, measuring
an “experienced mental event” ultimately relies on self-report. Indeed, histori-
cally intrusive thinking has been studied with self-report questionnaires (for an
overview, see Banich, this volume).
A general issue is how well the more objective, laboratory-based mea-
sures of intrusive thinking correlate with subjective measures. A common
finding for other constructs (e.g., impulsivity) is that they do not intercor-
relate particularly well (Nombela et al. 2014). Why might this be the case?
One notion is that the subjective measures are in some sense “noisier” and
more prone to error. It is well known, for instance, that people sometimes
have great difficulty in expressing their conscious evaluations or descrip-
tions of their thinking, and there may be considerable interindividual ability
in this capacity. However, this notion can perhaps be dismissed, as there
is also considerable evidence of superior test-retest reliability for question-
naires than objective measures, such as those based on reinforcement learn-
ing parameters (see Table 5 in Bland et al. 2016). It does appear likely that
objective tests may capture much narrower aspects of the construct under
study, thus resulting in a looser association with the more generic aspects
captured by a composite measure of a psychometrically well-designed ques-
tionnaire. Another important conclusion, however, is that the objective mea-
sures and the subjective responses obtained from questionnaires are simply
tapping into quite different processes, the latter most obviously addressing
the contents of monitoring operations in meta-consciousness. This reliance
on meta-awareness can be problematic and may be especially compromised
in clinical populations. More importantly, in the field of intrusive thinking,
there is a clear need for the development of better self-report instruments:
at present, there does not appear to be a measure that captures all aspects of
intrusive thinking in one instrument. In particular, there are four gaps in this
field (for discussion, see Banich, this volume).
1. Most questionnaires were designed with a particular clinical syndrome
in mind (e.g., PTSD, OCD, craving). As such, a general intrusive think-
ing scale does not exist.
2. For the most part, the questionnaires do not assess both the content
of the thought as well as the ability to control those thoughts, broken
down by content.
3. There are few, if any, questionnaires that specify the quality of intru-
sions across numerous dimensions, such as their form (verbal, images,
urges), vividness, salience, and frequency.
4. There are no questionnaires that systematically assess how intrusions
are triggered or the context in which people experience intrusions.
Looking forward, future research should focus on designing a new, theoreti-
cally motivated questionnaire, refining existing objective measures and capi-
talizing on the potential of computational approaches.
What Are the Neural Systems Relevant for

Intrusions and Their Control?
The exact neural underpinnings of intrusions are currently unknown, but we

could speculate about processes involved in intrusive thinking by combining
what we know from the study of cognitive control, including action control,
and the study of emotional memory processes.
Cognitive Control
Action Control
A critical prefrontal region for stopping movements is the right inferior frontal
cortex (rIFC) as shown by lesion and transcranial magnetic stimulation studies
(Aron et al. 2014). It is thought that the rIFC, in concert with the presupple-
mentary motor area, implements stopping through a (fast) hyperdirect pathway
to the subthalamic nucleus (STN) of the basal ganglia, and this suppresses tha-
lamic drive back to cortex, leading to movement cancellation (reviewed by Bari
and Robbins 2013; Jahanshahi et al. 2015; Wessel and Aron 2017). Importantly,
in the standard case, subjects apparently stop using a global mechanism. For
example, stopping the voice suppresses the hand (Badry et al. 2009; Cai et al.
2012). This broad skeletomotor suppression begins around 120 ms after the
Stop signal and lasts for 100 ms or more. It is thought that the broad suppres-
sion relates to a wider putative impact of the STN on basal ganglia output (Aron
2011), just as the degree of global motor suppression in the Stop-Signal para-
digm relates to the level of oscillatory power in the STN (Wessel et al. 2016).
Yet subjects can also stop selectively when they are forced to do so (Aron
2011). For example, they are able to stop one response (or hand) while continu-
ing with another with minimal interference, and this does not result in broad
motor suppression (Majid et al. 2012). It is possible that this form of selective
response suppression relates not to a hyperdirect cortico-STN connection but
rather to a frontal-striatal-pallidal, so-called, indirect pathway. For example, peo-
ple with degeneration of striatum and pallidum, who are thought to have an indi-
rect pathway disorder, could not stop as selectively in the paradigm and did not
show typical physiological signatures of selective stopping (Majid et al. 2013).
As will be discussed below, this global versus selective picture in motor
stopping may have relevance for global versus selective control over memory.
Future work could develop behavioral paradigms to look at this selective con-
trol on the analogy of selective response suppression, taking into account the
observation that selective stopping is best done when the subject proactively
sets it up ahead of time (Cai et al. 2011; Majid et al. 2012); that is, it could
be possible to go into a situation preparing to suppress a particular memory
intrusion.
Working Memory
The three working memory operations that we discussed above may be par-
ticularly relevant to intrusions:
1. Gain access or gate information into working memory.
2. Select information within working memory to be given priority.
3. Remove information from working memory.
In general, the neural correlates of working memory mechanisms have been
well described (D’Esposito and Postle 2015) and likely involve prefrontal ar-
eas that are involved in executive control processes (e.g., selecting among in-
formation in working memory), basal ganglia mechanisms that work to gate
information into working memory, and posterior brain regions that help to pro-
vide sensory or abstract (e.g., semantic) representations of information to be
activated and/or placed in working memory. Most of the paradigms utilized in
cognitive neuroscience require the confluence of processes whereby informa-
tion enters, is selected, and updated in working memory (e.g., the N-back task).
Nonetheless, the neural bases of some of these three main processes have been
distinguished.
With regard to gaining access to working memory, research suggests that

the basal ganglia may play an important role in determining when the gate
to working memory should be opened so as to let in new information (see
Badre, this volume). It is proposed that dopaminergic reinforcement learning
mechanisms help to provide information on when the gate to working memory
should be opened. Hence, this process is not conscious and controlled. The
basal ganglia form multiple loops with distinct regions of the cortex, which
are semi-segregated based on the nature of information they carry (e.g., sen-
sory, motoric). As such, these basal ganglia mechanisms might selectively or
concurrently act to allow access to working memory. In the case of intrusive
thoughts, especially when they are repetitive in nature, it may be that through
habitual learning, the intrusive content can more easily open the gate and gain
access to working memory. While alterations in the accessibility of informa-
tion to working memory may affect many different syndromes in which intru-
sive thoughts are observed, the specific basal ganglia loop involved may differ
according to the disorders (e.g., more motoric in substance use, more visual in
anxiety), as discussed by Balleine (this volume).
With regard to selecting information within working memory to be priori-
tized (or buffered from interference), processes often considered to be execu-
tive aspects of working memory, meta-analyses suggest that selection relies
on frontal and parietal regions across a variety of tasks (Nee et al. 2013). For
example, using Stroop and emotional Stroop tasks, research suggests that these
tasks tend to involve regions associated with the frontoparietal, cingulo-oper-
cular as well as dorsal and ventral attention networks (e.g., a meta-analysis by
Chen et al. 2018b), regions associated with cognitive control. Activation in
these areas has been found to be altered by characteristics related to intrusive
thought. For example, an individual’s degree of worry influences the degree of
activation (Engels et al. 2007) as well as the time-course of activation (Levin
Stilton et al. 2011) across lateral and medial prefrontal regions. Portions of the
superior parietal lobe may be particularly important in selecting or shifting
what information is currently within the attention focus of working memory
(Tamber-Rosenau et al. 2011). These regions are engaged in executive pro-
cesses more generally, not just specifically with regard to working memory.
With regard to removing information out of working memory, the picture is
less clear. As discussed by Banich (this volume), most methods examine removal
from the perspective of replacement; that is, how new information is placed into
working memory and/or how the current information is buffered from removal.
These operations involve cognitive control regions, including both the fronto-
parietal network and the cingulo-opercular network. Furthermore, variance in
activation in these areas has been observed in individuals, like those who are
depressed, who behaviorally show difficulty in removing certain types of infor-
mation (e.g., negative emotional information), as indexed by subsequent interfer-
ence from those items (Foland-Ross et al. 2013).
Less research has examined how one can take information currently held
within working memory and dispose of it completely. This has been a difficult
question to address because it is hard to verify that an item has indeed been
removed. Other than with behavioral methods such as self-report, which may
not be reliable, few (if any) methods are able to confirm that a thought has been
removed or inhibited. Yet understanding such processes could have important
implications for how interventions for intrusive thoughts might be created.
While there is little research in this area, some work has demonstrated that
brain imaging techniques can be used to help verify the removal of informa-
tion and simultaneously to examine the control mechanisms by which such
removal occurs. In one of the few studies of this nature, Banich et al. (2015)
utilized activity in the sensory cortex as a rough proxy for whether informa-
tion was currently active in working memory. With this approach, they verified
the presence of such activity when individuals were maintaining information
about a picture just viewed or a short tune just heard, or when they replaced
the item with something else. They also observed a lack of such activity when
the item was removed from current thought, both when participants specifi-
cally suppressed that item as well as when they did so by clearing their mind
of all thought. With regard to control mechanisms, a hierarchy of function was
observed. Common across the replace, suppress, and clear conditions—all of
which require a shift in attention away from the original item to something
else, compared to the maintain condition—was activation in superior pari-
etal regions implicated in shifting attention among items in working memory
(Koenigs et al. 2009). When information had to be removed from working
memory (the suppress and clear conditions) as compared to when an item was
present in working memory (the maintain and replace conditions), activation
was observed over regions of lateral prefrontal cortex involved in executive
processes that act on working memory. Finally, for the clear condition as com-
pared to all the other conditions, there was increased activation in the insula,
suggesting a shift of attention to bodily states (Craig 2011) as well as activa-
tion in inferior parietal cortex, which may represent a mechanism for altering
the bottom-up salience of information (Cabeza et al. 2008), either by reducing
the salience of visual information or by increasing salience of information de-
rived from bodily states (see also Fedota and Stein, this volume). In relation
to long-term memory, the global “clear” condition on the contents of working
memory may be analogous to a global stopping process that inhibits retrieval
of all information from the hippocampus, whereas the “suppress” condition
may be analogous to suppression of a specific memory, as discussed in more
detail below.
In summary, the brain regions relevant to working memory processes in-
volved in intrusive thought likely involve (a) content-specific cortical regions
needed to access the representation of information underlying the intrusive
thought: visual areas for visual images, language- or semantically related re-
gions for thoughts, limbic regions for emotional information, and hippocampus
for episodic memory, (b), basal ganglia mechanisms that influence what gains
access to working memory (see Badre, this volume), and (c) prefrontal mecha-
nisms involved in control processes that select and manipulate information
within working memory or act to engender its removal.
An unresolved issue concerns the extent to which the processes deployed
in suppressing contents actively held in working memory are distinct from
or overlapping with those involved in suppressing the long-term memory re-
trieval of unwanted thoughts, prompted by reminders (see section on retrieval
suppression), and whether these different types of mental control tap into
unique processes that may be differently affected in psychiatric disorders.
Long-Term Memory
Memory Encoding
For over a century, it has been recognized that memories are initially un-
stable, subject to interference, until they are stabilized through a process of
consolidation (Ribot 1882; Muller and Pilzecker 1900; Burnham 1903). Once
consolidated, long-term memories are largely protected from interference,
save windows of destabilization that may occur upon retrieval (Sara 2000;
Nader 2003). Reconsolidation involves a relatively brief (few hours) cascade
of molecular and cellular processes enhancing synaptic efficacy via structural
changes; a longer process (days and weeks) involves system-level connectivity
changes between the hippocampus and cortical areas (McGaugh 2000; Dudai
2012). Emotionally significant experiences trigger the release of adrenal stress
hormones, stimulating norepinephrine in the amygdala, which in turn modu-
lates plasticity in the hippocampus, cortex, and other brain regions (Cahill et
al. 1994; Southwick et al. 2002; Cahill and Alkire 2003; Strange and Dolan
2004; Hurlemann et al. 2005). Abundant evidence links dysfunction in these
circuits to psychological disorders in which intrusions are a symptom. Figure
9.3 presents a schematic overview of the neural circuits underlying these major
domains and their possible (dys)function in intrusive thinking.
To the degree that encoding of information is tied to a specific event, such
as that often associated with PTSD, aberrant functioning in the emotion cir-
cuits as well as the hippocampal and medial temporal structures that support
episodic memory are likely involved (Liberzon and Martis 2006; Pitman et al.
2012). Evidence points to hyperactivity in the amygdala and dorsal anterior
cingulate cortex (the putative human homologue of the prelimbic medial pre-
frontal cortex), whereas the ventromedial prefrontal cortex and hippocampus
evince hypofunction, accompanied by impaired extinction learning and recall
(Milad et al. 2009; Milad and Quirk 2012; Logue et al. 2018). Most recently,
specific computations of threat, including learning parameters such as value,
prediction error, and learning rate, have been characterized in PTSD. It was
found that PTSD severity was related to overweighing of prediction errors
dlPFC vmPFC
Cognitive
regulation
Retrieval
suppression
Inhibitory control
Episodic (e.g., extinction)
memory
Hippocampus Context
Sensory input Amygdala Striatum
Emotional learning Goal-directed behavior

(e.g., threat conditioning) (e.g., active avoidance)
Hypothalamus/brain stem
Emotional reaction
(e.g., freezing, sweating)
Figure 9.3 Most of what we know about the neural mechanisms of emotional learning
and memory comes from animal studies that utilize threat conditioning and extinction
as a model. Based on these studies, sensory information from neutral stimuli in the en-
vironment that reliably coincide with emotionally significant outcomes converge in the
lateral nucleus of the amygdala where associative learning occurs, conferring emotional
value on the conditioned cues. Projections from the lateral to the central nucleus engage
descending projections to the hypothalamus and brain stem, which mediate the expres-
sion of the conditioned response (e.g., freezing). Projections from the lateral to the basal
nucleus onto the striatum form a path that promotes active coping and goal-directed
behavior. The prelimbic region of the medial prefrontal cortex (mPFC) connects with
the amygdala to sustain the expression of emotional responses, while the infralimbic
region acts to counteract amygdala output and diminish emotional responses. As such,
the infralimbic mPFC (ventral mPFC in humans) is a critical region in the acquisition
and recall of extinction learning. The hippocampus exerts contextual control over the
expression of threat learning. The dorsolateral prefrontal cortex (dlPFC) is the region
involved in top-down cognitive regulation of emotional responses by influencing amyg-
dala via the vmPFC (Hartley and Phelps 2010; Schiller et al. 2010; Milad and Quirk
2012). The dlPFC also exerts top-down modulation of hippocampal activity to induce
retrieval suppression (Anderson et al. 2016).
(akin to enhanced sensitivity to negative surprise) and to impaired amygdala

and striatal tracking of the negative value of conditioned cues (Homan et al.
2019). Neuroimaging studies that examine the relation between viewing ana-
log trauma and subsequent intrusions are sparse but do tentatively suggest that
activation of regions in the salience network distinguishes distressing scenes
that later intrude compared to equally distressing scenes that do not (Bourne et
al. 2013; Battaglini et al. 2016; Clark et al. 2016), highlighting that consolida-
tion processes are important in the formation of intrusive memories. Together,
these studies invoke neural mechanisms for the learning and flexible modula-
tion of emotional memories. Aberrant functioning of the salience and memory
circuitries might induce emotional inflexibility, lack of proper contextual con-
trol, and impaired ability to diminish inappropriate emotional reactions, allow-
ing strong negative memories to persist and intrude.
Similar mechanisms may come into play if a memory is associated with
particular actions or procedures, such as may occur in substance use disorders.
Here, the encoding likely involves basal ganglia systems and alterations in
specific cortical processes (Balleine, this volume), such as visual regions that
may become tuned to particular stimuli (e.g., paraphernalia associated with
substance use). Across disorders, when information is associated with emo-
tional salience or significance (e.g., salience of negative emotional information
in depression, association of loud sounds with a traumatic event in PTSD),
such encoding likely involves the amygdala, as well as the striatum, insula,
and dorsal anterior cingulate cortex (Fedota and Stein, this volume). However,
whether intrusive memories are stored in a qualitatively different way than
other emotional memories, or are merely strong memories and therefore more
accessible, is still a topic of debate (Berntsen and Rubin 2008, 2013; Brewin
2014; Bisby and Burgess 2017).
Retrieval Suppression
Although we know little about the neural mechanisms of intrusions, we know

a great deal about the inhibitory control processes deployed to reduce their
occurrence. For an in-depth review of the neural basis of memory control,
see Banich et al. (2009) and Anderson and Hanslmayr (2014). For a review
of memory control in relation to emotion regulation, see Engen and Anderson
(2018). For a detailed consideration of memory control in relation to primate
anatomical pathways and neural circuits, see Anderson et al. (2016).
Prefrontal-Hippocampal Interactions as a Basis for Retrieval Suppression.

Research on retrieval suppression was initially premised on the parallel be-
tween stopping prepotent actions in response to triggering stimuli and stopping
internal processes, such as memory retrieval (in response to reminders), to
control intrusive memories and thoughts (Anderson and Green 2001). In both
cases, a stimulus (or, sometimes in the case of retrieval stopping, an internally
generated cue) initiates an automatic process (action preparation or retrieval)
that the person wishes to stop, and both trigger a race between a Go and a Stop
process. Given this analogy, one might expect both similarities and differences
between retrieval and action stopping: similarities in the prefrontal control
regions engaged in service of stopping, but differences in the target regions
with which control processes interact to implement stopping. Because retrieval
stopping involves stopping retrieval and not physical actions, a plausible sup-
pression target would be the hippocampus, a brain structure involved not only
in encoding new memories but also in their retrieval, at least for recently ac-
quired events and thoughts (Anderson and Green 2001; Anderson et al. 2004;
Anderson and Hanslmayr 2014).
Abundant evidence indicates that retrieval suppression, unlike motor re-
sponse inhibition, relies on prefrontally mediated downregulation of activity in
the hippocampus and other medial temporal lobe regions to stop retrieval, pre-
sumably by preventing pattern completion (Anderson et al. 2004, 2016; Depue
et al. 2007; Banich et al. 2009; Levy and Anderson 2012; Gagnepain et al. 2014,
2017; Benoit et al. 2015; Schmitz et al. 2017). Suppression reduces hippocam-
pal activation not merely relative to active retrieval in the Think condition, but
also relative to passive baseline conditions, and this negative BOLD response
arises from negative coupling between the right lateral prefrontal cortex and
the hippocampus, established using effective connectivity analysis (Benoit and
Anderson 2012; Gagnepain et al. 2014, 2017; Benoit et al. 2015; Schmitz et
al. 2017). Indeed, within-subject comparisons of action and memory stopping
establish a clear double dissociation, with retrieval suppression downregulating
the hippocampus more than action stopping, but action stopping downregulat-
ing motor cortical areas (M1) more than retrieval suppression (Schmitz et al.
2017). Stopping of unwanted memories and thoughts thus involves a distinct
frontohippocampal inhibitory control pathway that suppresses hippocampal
activity. Hippocampal suppression appears to be a blunt instrument that acts
globally on the hippocampal state. For example, when a person tries to suppress
retrieval to depress awareness of a particular unwanted thought, the forgetting
arising from that suppression is not limited to the item people intend to suppress;
rather, any other recently encoded memories that occur either before or after the
act of suppressing something are also forgotten, even if they are entirely unre-
lated to the content being suppressed (Hulbert et al. 2016). Thus, suppression
of unwanted thoughts induces an amnesic shadow in the temporal surround of
the suppression attempt, creating both anterograde and retrograde amnesia ef-
fects in healthy people. This finding has been linked to the global suppression
of hippocampal processes that not only stop retrieval but also disrupt encoding
and stabilization processes necessary to retain recent experiences (Hulbert et al.
2016). This global, systemic disruption of hippocampal activity is analogous to
the global stopping identified in motor response inhibition.
What do we know about how the prefrontal cortex achieves this form of
inhibitory control over hippocampal activity? Primate anatomical studies tell
us that top-down suppression of hippocampal activity is unlikely to be direct,
not only because there are no direct connections between the lateral prefrontal
cortex and the hippocampus (Anderson et al. 2016), but also because long-
range projections from the hippocampus are largely excitatory. To achieve an
inhibitory effect in the hippocampus, if the negative BOLD response is truly
inhibitory, the prefrontal cortex must drive local populations of inhibitory
interneurons within the hippocampus to disrupt its function. Because all inter-
neurons in the hippocampus are GABAergic, this observation suggests that
individuals with higher concentrations of hippocampal GABA may show a
superior ability to suppress hippocampal activity by prefrontal influence.
Recently, Schmitz et al. (2017) found evidence of this, using a multimodal
imaging study that combined fMRI with magnetic resonance spectroscopy.
These findings indicate that people with higher concentrations of hippocam-
pal GABA showed greater downregulation during retrieval suppression, more
successful forgetting of intruding thoughts, and greater negative coupling be-
tween the right prefrontal cortex and the hippocampus. As such, local concen-
trations of hippocampal GABA may provide a pivotal function that enables the
prefrontal cortex to implement long-range inhibitory influence necessary for
control over intrusive thoughts. This discovery sheds new light on evidence of
diminished hippocampal GABA in many disorders characterized by intrusive
thoughts (Schmitz et al. 2017), which may be a heretofore unrecognized risk
factor in the pathogenesis of disordered control over intrusive thoughts.
Despite the unique hippocampal targets involved in implementing re-
trieval suppression, it is equally clear that both retrieval and action stopping
processes engage overlapping regions in the dorsolateral (BA 9/46/10) and
ventrolateral prefrontal (BA 44/45) cortex, suggesting the existence of do-
main general supramodal inhibitory control regions that may dynamically
recouple with task-specific target regions (Depue et al. 2016; Schmitz et al.
2017; Guo et al. 2018). These domain general regions are strikingly right
lateralized, strongly consistent with the long-standing claim by Aron et al.
(2004, 2014) that inhibitory control is right lateralized, although the regions
clearly include both dorsolateral and ventrolateral regions, and not simply
ventrolateral prefrontal cortex. These supramodal regions have been identi-
fied both through within-subjects conjunction analyses of action and retrieval
stopping (Depue et al. 2016; Schmitz et al. 2017) as well as conjunctions
performed on quantitative meta-analyses of independent studies from many
laboratories (Guo et al. 2018). Interestingly, action stopping and retrieval
stopping also appear to engage highly colocalized regions within the basal
ganglia (Anderson et al. 2016); for a detailed discussion of anatomical hy-
potheses, see Depue (2012), Guo et al. (2018), Paz-Alonso et al. (2013), and
Balleine (this volume).
Cortical Modulation and the Reinstatement Principle. Although the discus-

sion of direct retrieval suppression emphasized hippocampal suppression as
the principal mechanism through which intrusive memories and thoughts are
controlled, the hippocampus is not the only target of inhibitory control dur-
ing suppression (see Banich as well as Balleine, this volume). For example,
when people suppress visual objects or scenes, downregulation takes place
in the hippocampus as well as in visual cortical regions, such as the fusiform
cortex and the parahippocampal place area, respectively. This leads to the
generalization that areas outside the hippocampus involved in reinstating the

unwanted memory in awareness are suppressed in parallel with the hippocam-
pus by the right lateral prefrontal cortex, an idea known as the reinstatement
principle (Gagnepain et al. 2014; Gagnepain et al. 2017).
Intrusive Thoughts and the Triggering of Control. Intrusions of unwanted

contents into awareness during retrieval suppression are particularly important
for triggering top-down inhibitory control processes that suppress the hippo-
campus and cortical regions. Using the intrusion judgment procedure outlined
earlier, Levy and Anderson (2012) found that downregulation of activity in
the hippocampus and other medial-temporal lobe regions was largely con-
fined to trials in which participants reported that a memory had intruded into
consciousness. Strikingly, the extent of hippocampal downregulation during
intrusions predicted subsequent forgetting with a correlation of 0.7, whereas
hippocampal activity during non-intrusions was unrelated to later forgetting.
These findings suggest that intrusions play an important role in triggering top-
down control by the prefrontal cortex to cancel the retrieval process, and that
the critical inhibitory action that disrupts later retention of the intrusive thought
arises in the purging of the intrusion from awareness. Consistent with this pos-
sibility, Gagnepain et al. (2017) replicated Levy and Anderson’s (2012) evi-
dence for intrusion-specific downregulation in the hippocampus with aversive
scenes and showed that negative coupling between the right dorsolateral pre-
frontal cortex and the hippocampus was significantly greater during intrusions
than during non-intrusions.
The findings by Gagnepain et al. (2017) are especially relevant to the current
discussion, given that they involve the suppression of unpleasant and intrusive
images, prevalent in psychiatric disorders. Gagnepain et al. (2017) also found
intrusion-specific downregulation in both the amygdala and the parahippocam-
pus. These downregulations robustly predicted both intrusion frequency and
later reductions in negative affect for the suppressed content: the stronger the
downregulation in the amygdala, the lower the number of intrusions and the
greater the subsequent reduction in negative affect for the suppressed scene.
These findings indicate that top-down inhibition during intrusive thoughts
plays a critical role in modifying the representations that support both memory
and emotion about the offending content. Consistent with this, Legrand et al.
(2019) found that suppressing unpleasant images from awareness also signifi-
cantly reduced later psychophysiological measures of emotion elicited by the
scenes, such as heart rate deceleration. Similar findings have now been ob-
served with skin conductance responses (Harrington et al. 2020).
The importance of purging intruding thoughts from awareness highlighted
here points back to the discussion above of what relationship, if any, the
processes involved in stopping retrieval have to those involved in regulat-
ing working memory. One assumes that when an intrusion is retrieved and
enters awareness, there is a good chance that the intruding content has entered
working memory, however briefly. Some evidence supports this. For exam-
ple, Hellerstedt et al. (2016), using event-related potentials, found evidence
that the frontal negative slow wave (NSW), observed over the right prefron-
tal cortex, is modulated during intrusions. This component has been linked
in prior work to the storage of information in working memory. Consistent
with this, participants during Think trials show a prolonged NSW that lasts
throughout the full several seconds of the trial; in contrast, non-intrusions
show little evidence of this, consistent with the exclusion of items from work-
ing memory. Intrusions, however, showed a brief increase in the NSW, which
was rapidly eliminated within the first seconds of the trial, suggesting a brief
penetration of the intruding item into working memory. Perhaps relatedly,
Castiglione et al. (2019), using time frequency analysis, found evidence that
during No-Think trials, there is a robust increase in frontal beta component
during non-intrusions. Given the prior linkage of this component to motor
response inhibition, these findings suggest that intrusions may reflect an ini-
tial failure of inhibitory control that allows the intruding content to penetrate
working memory.
Related Phenomena Observed with Item-Method Directed Forgetting. Al-

though we have focused primarily on retrieval suppression, related work
on item-method directed forgetting supports the hypothesis that, in parallel
with retrieval suppression, people can also suppress encoding. By encod-
ing suppression, we mean the possibility that the same inhibitory control
processes which modulate hippocampal activity during memory retrieval
to disrupt retention may also be deployed shortly after encoding to termi-
nate stabilization processes in the hippocampus that might promote the for-
mation of an enduring memory. For example, several studies indicate that
when participants are instructed to forget an item that they just encoded into
memory on the preceding trial, activation increases in the right dorsolateral
prefrontal cortex and decreases in the hippocampus. As with retrieval sup-
pression, connectivity analyses indicate that the prefrontal cortex couples
with the hippocampus during Forget trials, especially on trials when the item
is successfully forgotten (Rizio and Dennis 2013; Wierzba et al. 2018). In
another compelling study, intracranial recordings indicated that lateral pre-
frontal cortex interacts with the hippocampus during instructions to forget to
promote forgetting (Oehrn et al. 2018). These findings converge to suggest
that suppressive processes are not limited to controlling intrusive retrievals,
but may be also be used prophylactically shortly after an unpleasant experi-
ence to limit the footprint of that experience in memory. Moreover, this ap-
plication of inhibitory control to suppress unwanted contents immediately
after they are encountered seems related to inhibitory processes involved
in purging the contents of working memory (e.g., Holmes et al. as well as
Banich, this volume), although these two strands of research are not usually
considered together.
At present, little is known about the neural underpinnings of intrusive think-

ing, and the underpinnings presumably vary across its different manifesta-
tions. In contrast, research on cognitive control and emotional memory has
identified key systems involved in generating and suppressing intrusions.
Concerning intrusion generation, areas involved in signaling salience, such as
the amygdala and striatum, not only play a role during the encoding of salient
events or thoughts, but also respond to reminders of those events (external
cues). Their salience signal may help bring back to mind memories of these
events or thoughts, possibly via reinstatement of multimodal cortical repre-
sentations of these events. In addition, these areas may themselves generate
internal cues (mood states) that trigger intrusions. With regard to the suppres-
sion of intrusions, frontal areas implicated in inhibitory control over actions
and thoughts play a clear role. In particular, the dorsolateral prefrontal cortex
can suppress the reinstatement of a memory or of an imagined future event by
downregulating hippocampal and neocortical activity, while also downregu-
lating emotional responses via amygdala suppression. Suppression reduces
the frequency of intrusions and impairs memory, also for new information
presented around the time of retrieval suppression (i.e., creating an amnesic
shadow). Some evidence indicates that it also reduces negative affect associ-
ated with suppressed content.
What Are the Implications of Intrusive Thinking?
In this chapter we have revisited the definition of intrusive thinking by sys-

tematically considering all the circumstances in which intrusions might occur
and their manifestations across health and disorders. We define intrusions as
being interruptive, salient, experienced mental events and propose that clinical
intrusive thinking differs from its nonclinical form with regard to frequency,
intensity, and maladaptive reappraisal. We have reviewed the neurocognitive
processes underlying intrusive thinking and their control, including action con-
trol, working memory processes, long-term memory encoding, retrieval and
suppression, and methods for studying them.
Functional Perspective: The Adaptive Nature of Intrusive

Thinking and the Desirability of Suppression
Despite being commonly associated with mental health disorders, intrusive

thinking commonly occurs in the absence of psychological problems and is
thus by itself not indicative of pathology. In fact, there are many instances in
which having a thought pop into mind to disrupt ongoing cognitive processes
is actually beneficial, such as recalling an action that is required (e.g., paying
a bill) or solving a problem (an “aha” moment) (see also Monfils and Buss,
this volume). If mind wandering is a form of intrusive thinking, research
shows that the capacity to “jump out of the present set” can be adaptive and
relates to creativity. Its adaptiveness is likely reduced when mind wandering
becomes excessive or too diffuse, such as in ADHD, or when the content is
unpleasant.
The potential adaptive nature of intrusive thinking may be independent
of the desirability to control it. In healthy individuals, frequent and excit-
ing thoughts (e.g., love infatuations) might disrupt concentration on the
task at hand and therefore not be adaptive, yet still wanted in the moment.
Alternatively, intrusive thinking about mistakes that we have made may allow
us to adjust our behavior and become a better person, yet unwanted in the
moment. The discussion below on sociocultural implications provides com-
pelling examples of this. In addition, as a clinical symptom, intrusive think-
ing is not always unwanted (e.g., exciting flashforward thoughts that occur
during a manic episode), yet often maladaptive. Whether the opposite is also
possible (i.e., unwanted intrusive thinking in mental health disorders may in
some cases serve an adaptive purpose, such as in acute stress disorder as part
of processing trauma), is an open question that requires more research.
Clinical and Therapeutic Implications
Could excessive intrusive thinking be an endophenotype (or neuroendophe-

notype) for mental disorders? According to the novel, more dimensional ap-
proach to understanding psychiatric nosology (Cuthbert and Insel 2013), the
propensity toward intrusive thinking could underlie several otherwise distinct,
categorically defined disorders, such as OCD and depression, the implications
being comorbid, shared dysfunctions of common neural systems or networks,
with obvious implications for treatment. Considering some of the dimensions
already postulated, there are clear relationships with such constructs as in-
hibitory control and working memory. It remains unclear, however, whether
intrusive thinking comprises a unitary construct itself or is a collection of
phenomena that meet our definitions. For example, intrusive thinking could
arise as an emergent feature of different neural networks processing perceptual
inputs, on one hand, or neural networks underlying internal factors such as
intentions and mood states, on the other. However, such malfunctions in dif-
ferent networks could depend, for example, on a common molecular or neu-
rotransmitter deficit. This issue can perhaps best be resolved when we have
firmer information about the neural substrates of intrusive thinking. There are
some promising indications of this (see chapters by Philips, Fedota and Stein,
Balleine, Badre, Gourley et al., and Roberts et al., this volume) as well as pos-
sible genetic relationships, which will depend on having precise definitions of
the phenotype obtained, for example, through definitive objective tests and a
standardized general instrument for assessing intrusive thinking.
The mechanisms and paradigms discussed have implications for therapeutic

applications. An idea often discussed in the clinical literature is that suppress-
ing intrusive thoughts in clinical disorders is counterproductive, an idea sub-
stantially influenced by Wegner’s white bear thought suppression procedure
(Wegner et al. 1987; Wegner 1997). However, substantial recent evidence with
the Think/No-Think procedure suggests that findings from the white bear para-
digm have been overgeneralized and that other processes, such as retrieval sup-
pression, are effective in modulating thoughts, at least in healthy individuals.
If so, there may be situations in which suppressing thoughts could be useful
to improve an individual’s ability to function, even in clinical disorders. For
example, it may be that both trauma-focused interventions, such as eye move-
ment desensitization reprocessing, and paradigms using cognitive interference,
such as memory retrieval procedure and visuospatial task (Holmes et al., this
volume), which appear to be effective in reducing the frequency of intrusive
memories of trauma, work through a process similar to the retrieval-suppres-
sion mechanism described above. While it is clear that in some disorders (e.g.,
OCD), attempts to suppress intrusive thoughts can lead to rebound, clinical
disorders involving intrusive thoughts are heterogeneous, as are patient popu-
lations who have these disorders. Thus, it is critical to explore the therapeutic
implications of paradigms which have demonstrated efficacy in controlling,
replacing, and suppressing intrusive thoughts in human laboratory settings (see
also Brewer et al., this volume).
A particularly promising therapeutic approach for fostering the effective
control of intrusive thoughts is mindfulness-based cognitive therapy, which
integrates meditation techniques with cognitive behavioral strategies (Külz et
al. 2014). One of the primary skills taught in such approaches is the ability to
control one’s thoughts to focus on breathing. The act of releasing thoughts in
mindfulness practices is highly reminiscent of the control strategies invoked
in the Think/No-Think paradigm, further illustrating their potential pertinence
to the treatment of intrusive thoughts. Mindfulness practices may also be help-
ful in furthering individuals’ meta-awareness of having intrusive thoughts
in the first place (Baird et al. 2014). Specifically, such practices may enable
individuals to identify episodes of unwanted thoughts that might have oth-
erwise been experienced but evaded explicit acknowledgment (Baird et al.
2013a; Takarangi et al. 2014). Identifying intrusive thoughts in the light of
meta-awareness may enable individuals to invoke the necessary mental control
strategies required to release them.
Sociocultural Context
Intrusive thinking and its control (or lack thereof) take place within a sociocul-
tural context. Examples of this include the AIDS-HIV epidemic of the 1980s.
In his book, The Man Who Couldn’t Stop: OCD and the True Story of a Life
Lost in Thought, David Adam (2014) provides a vivid example of how frequent
concerns about infection drove the induction of his own obsessive and com-
pulsive symptoms. Certainly, there are many other examples of widespread
general concern (e.g., nuclear war, Brexit, global pandemics) that intrude into
our everyday consciousness and may lead to pathological consequences.
Although many of us successfully resist such concerns, this too can lead
to societal consequences, as can be currently observed through the striking
example of the climate crisis. Since the Industrial Revolution, global tempera-
tures have increased by 1°C and humankind is well on track to experiencing an
additional increase of 0.5°C by 2030 (Xu et al. 2018). In line with the worst-
case scenario put forth by the Intergovernmental Panel on Climate Change
(IPCC 2018), this half a degree will likely correspond to a 50% increase in
extreme weather events worldwide (e.g., droughts, floods, snowstorms, hur-
ricanes, cyclones) and exact devastating consequences, including mass migra-
tion, agricultural failure, and deadly heat events (Xu et al. 2018). The near total
consensus of climate scientists, exemplified by the Paris Agreement, which
almost all governments signed (and almost none are honoring), is that we have
to reduce emissions soon. Failure to do so will result in continued temperature
increases that will soon be beyond human control and ultimately lead to a “hot-
house planet” by 2100, or perhaps even sooner (Wallace-Wells 2019).
The climate crisis is surely creating daily thought intrusions in hundreds
of millions of people. Such intrusions are likely characterized by interrupting,
salient, experienced events (imagery, emotion, moral feeling) that recur. This
example illustrates the complexity of judging whether intrusive thinking is mal-
adaptive or not, and the range of responses that people can have. For instance,
against the unimpeachable backdrop of scientific knowledge, such intrusions
appear highly adaptive: they compel action and yet only a very small minority
of people are currently engaged in action—the great majority of global citizens
are not taking action. Of these, some may deny the science or the predicted
impacts, or they may accept the science and impacts but deny that any serious
action is warranted (e.g., because they are ideologically committed to the cur-
rent economic system). In Western societies, polling shows that most people
fall in the latter category (accepting the science and the impacts), yet they are
not acting beyond some minor adjustments in their personal lives, possibly be-
cause action would be inconvenient to one’s career or lifestyle or because it
would require confronting grief and fear in a way that one is not yet prepared
to do. Since such people do understand and accept the terrifying imperative to
act, but are not doing so, they could be characterized as exerting control over
the intrusive thinking caused by the climate crisis. The form of control being
used is probably based on reappraisal: people express degrees of fatalism (“the
problem is too big or too hopeless,” “it is too late”), nihilism (“humans deserve
what is coming to them”), a deferral of responsibility to policy makers (“it is
a government problem”), or presently unwarranted faith in technology alone
to solve the problem (Hansen 2018). Nonetheless, as more extreme weather
events, for instance, increase over time, so too will levels of anxiety, until the
intrusion breaks through the threshold of control and people clamor for action.
In summary and in accordance with the definition of intrusive thinking as
interrupting, salient, experienced events (imagery, emotion, moral feeling) that
are also recurrent, the climate crisis is clearly generating intrusive thinking in
a wide range of people around the world. This, in turn, creates conflicts which
can inculcate attempts at control (e.g., at the reappraisal level). Such tensions
exacerbate poor mental health as well as the feelings of instability and fear,
which are already driving populist political regimes (Latour 2018). Meanwhile
fake news disseminated within the broader system functions to degrade the
salience signal that is currently essential to generate the type of population-
level intrusive thinking that would compel quick action. This sociocultural
example is interesting because it represents the flip side of many of the psy-
chiatric-related examples discussed in this chapter. Whereas intrusions were
often characterized as pathological, the climate crisis example demonstrates
how intrusive thinking is a good thing for our biosphere and civilization, and
attempts to control these intrusions are maladaptive. The survival of human
civilization depends on more intrusive thinking right now.
Concluding Summary
Our discussion focused on the psychological bases of intrusive thinking and

its control that may occur every day in healthy individuals, as well as in
psychiatric disorders. We surveyed the range of phenomena that can be con-
strued as examples of intrusive thinking and endeavored to reach satisfactory
definitions and classifications of the different forms of intrusive thinking that
will aid further research. The least constrained definition was one empha-
sizing the interruptive, salient, and experienced nature of intrusive mental
events (as compared with a common definition which specifies unwanted
and conscious, as well as interruptive criteria). Recurrence is a further prop-
erty which may be more important for psychiatric manifestations. Agency,
meta-consciousness, and (mal)adaptiveness or desirability are considerations
that further define the boundaries of what can be considered as intrusive
thinking. Based on this analysis, PTSD and OCD appear to be prototypi-
cal disorders of intrusive thinking, although its elements appear in a range
of other psychiatric diagnoses, including addiction and depression, though
probably not psychosis.
The main part of our discussion focused on neurocognitive mechanisms
of intrusive thinking and its control, which included impairments particularly
in the regulation of memory retrieval, as well as of affect and action control.
We brainstormed future possible approaches to investigating intrusive think-
ing, with priorities for designing a new, theoretically motivated question-
naire, refining existing objective measures and capitalizing on the potential of
computational approaches. Finally, we considered the importance of this in the

context of broader social-cultural and clinical-therapeutic issues.
10
Networks Relevant to
Psychopathology and
Intrusive Thought
John R. Fedota and Elliot A. Stein
Abstract
Intrusive thoughts are regular occurrences in healthy cognition. Across a variety of
psychiatric conditions, however, such thoughts can become unconstructive and per-
severative. Failures in computations to estimate the salience of the content of these
thoughts are at least partly responsible for these clinically relevant disease symptoms.
This chapter reviews neuroimaging results that show specific and related dysfunction
in the calculation of salience at multiple neuroanatomically and functionally linked
regions of interest, both cortically and subcortically. Transdiagnostic evidence for dys-
function in the striatum, thalamus, and prefrontal cortex is reviewed, as is a theoretical
framework placing these regional findings in the context of large-scale brain networks.
It is argued that changes in nodal function and network communication are signatures
of a failure to properly shape predictions about the reliability and utility of external
and internal stimuli, leading to maladaptive attentional capture and behavior, including
intrusive thoughts.
Introduction
Spontaneous thought that is unrelated to current task demands is a normal part

of healthy cognition. According to estimates from thought probe experiments,
30–50% of waking cognition is unrelated to specific tasks (Kane et al. 2007;
Killingsworth and Gilbert 2010). However, even in normal cognition, this
frequent mind wandering appears to bear an emotional cost. When prompted
during mind wandering, participants reported being less happy than during
task-focused cognition (Killingsworth and Gilbert 2010).
Spontaneous thoughts rarely occur only once. Instead, they are often recur-
rent. Across all types of repetitive thought, a common feature is an internal
focus on “one’s self and one’s world” (Segerstrom et al. 2003). The content
188 J. R. Fedota and E. A. Stein
of these internally focused thoughts can be either positive or negative, such

as daydreaming versus worry (Watkins 2008; Christoff et al. 2016). An exten-
sive taxonomy of repetitive thoughts exists (Watkins 2008), which describes
both constructive and unconstructive consequences of intrusive and repetitive
thoughts that can be linked to psychiatric conditions.
The ubiquity and diversity of repetitive thoughts during cognition beg the
question: What characteristics mark the transition from healthy, spontaneous
thought to clinically relevant, repetitive intrusive thought? It appears that the
valence of the repetitive thought is an important factor in determining clini-
cal relevance. In Watkins’s (2008) taxonomy, negatively valenced thoughts
with unconstructive consequences include depressive rumination, worry, and
perseverative cognition. These consequences can clearly be linked to psychi-
atric conditions such as anxiety disorders, major depressive disorder (MDD),
obsessive-compulsive disorder (OCD), posttraumatic stress disorder (PTSD),
substance use disorders (SUDs), and schizophrenia (SZ). Negative repetitive
thoughts (NRTs) become clinically relevant when their magnitude or fre-
quency increases or when they become perseverative and difficult to control or
eliminate (Kalivas and Kalivas 2016).
Cognitive Constructs Relevant to Negative Repetitive Thoughts
Beyond a taxonomy of NRTs, a description of the cognitive construct(s) under-

lying these clinical phenomena is necessary. Here, the goal is to operationalize
the definition of NRTs as an intermediate step to identifying specific brain
regions, circuits, and large-scale networks where the identified constructs are
neurobiologically instantiated.
Cognition, intrusive or not, can be conceived as the making of predictions
about the environment, testing those hypotheses via sensory processing, and
subsequently updating and refining these predictions based on experience.
Thus, optimal interaction with the world requires accurate beliefs about the
environment and the ability to update those beliefs as new reliable evidence
is encountered. This process can be computationally formalized as Bayesian
predictive coding, a general theory of brain function that specifies how goal-
directed behavior is motivated through the integration of prior beliefs with sen-
sory information (Rao and Ballard 1999; Doya et al. 2007; Itti and Baldi 2009;
Friston et al. 2012; Aitchison and Lengyel 2017). This framework integrates
disparate cognitive processes including learning, reward, executive control,
attention, and sensory processing.
Briefly, Bayesian predictive coding involves the integration of prior be-
liefs and available sensory evidence to refine posterior estimates of beliefs.
Any mismatch between these two distributions signals the need to update
beliefs considering the evidence encountered. The computational weight
given to probabilistic estimates of either prior beliefs or sensory evidence is
Networks Relevant to Psychopathology and Intrusive Thought 189
governed by the relative precision of each (Mathys et al. 2011). That is, pre-
cisely estimated priors are less susceptible to modulation based on sensory
evidence, while poorly estimated (imprecise) priors are readily adjusted by
sensory evidence. As different sensory inputs can vary in their precision and
information content, one cognitive imperative is to estimate which available
sensory input will provide reliable and informative information to calculate
better posterior estimates of the environment (Parr and Friston 2017, 2019).
To this end, unambiguous sensory data should be amplified when present and
sought when absent.
Cognitively, this predictive estimation of potentially reliable sensory
sources has been described as the attribution of salience (Parr and Friston
2019). By salience, we mean a quality that is particularly noticeable or deemed
important to the individual in a given context (Uddin 2014; Kahnt and Tobler
2017; Miyata 2019). As such, elements that provide unambiguous sensations
should be ascribed higher salience, as they will provide reliable information
for the adjustment of posterior estimates and iteratively improve subsequent
decisions. The degree to which new information alters posterior estimates as
compared to prior beliefs is termed Bayesian surprise (Itti and Baldi 2009;
Barto et al. 2013).
Salience is closely related to value, but a key distinction is that value is a
signed currency that varies monotonically from negative to positive whereas
salience is an unsigned currency, where both negative and positive predicted
outcomes can have equivalent salience (Kahnt and Tobler 2017). This is be-
cause both positively and negatively valent elements of the environment can
improve posterior estimates; each can be informative in the refinement of pos-
terior estimates.
When working properly, this iterative cycle of hypothesis (prior)–result
(sensory evidence)–conclusion (posterior) allows for the flexibility and learn-
ing characteristics of human cognition. However, in the case of NRTs, im-
proper predictions of the salience of elements in the environment will lead
to suboptimal processing, including perseveration on elements with overly
precise priors and/or a failure to guide attention to elements with weak priors.
For example, if previous experience creates an overly precise prior belief
about the reliability of information to be gleaned from a stimulus (e.g., a drug cue
or emotionally valent memory), its salience will be increased. Such an increase in
anticipated salience will lead to the focus of greater attentional resources on the
stimulus, potentially at the expense of alternatives in the environment. This dys-
regulated focus on one thought at the expense of others is a hallmark of intrusive
thought, as described in the following sections. Indeed, the computational frame-
work of Bayesian predictive coding has been shown to be useful in describing
specific deficits across a variety of psychiatric conditions associated with NRTs:
hallucinations in SZ (Sterzer et al. 2018), drug cravings in SUD (Gu and Filbey
2017), and perseverative focus in OCD (Levy 2018).
Below, we evaluate neuroimaging data across nodes and networks previ-

ously implicated in the cognitive processes related to the estimation of prior
probabilities (salience) and attentional modulation within the framework of
Bayesian predictive coding. Not surprisingly, the multiple points of failure
in the information processing cascade requires discussion of a wide range of
implicated brain regions. In addition, these individual brain regions can be
supra-ordinately organized as nodes in large-scale networks, providing a sys-
tems-level perspective on dysfunctional communication associated with the
estimation of salience. Each source of potential salience attribution dysfunc-
tion will be addressed in turn.
Ventral Striatum and Potentiated Response
A large body of literature has shown the striatum to be responsive not only to
value and reward, but also to the salience of a given stimulus, independent of
its positive or negative valence. In humans, these reward (magnitude * signed
valence) and salience (magnitude * absolute value of valence) responses are
separable: salience-evoked activation is seen in the ventral striatum (Zink et al.
2006; Jensen et al. 2007; Bartra et al. 2013), the insula, and the dorsal anterior
cingulate cortex (dACC), while reward (positive valence) is also encoded in
the striatum and across various brain regions, including the orbitofrontal cortex
(OFC) (Litt et al. 2010; Bartra et al. 2013; Kahnt et al. 2014).
Dysfunction in striatal signaling related to the identification of salient
environmental stimuli is observed across psychiatric conditions associated
with NRTs. When compared to healthy individuals, those at risk for devel-
oping psychosis show heightened activation to task-irrelevant stimuli within
the ventral striatum (Roiser et al. 2012; Schmidt et al. 2016). The observed
increase in ventral striatal activation to irrelevant stimuli suggests an over-
sensitivity to uninformative cues in the environment. Within the Bayesian
predictive coding framework described above, during normal cognition these
irrelevant cues should be ascribed reduced salience, as they provide little to
no sensory information with which to refine posterior estimates of the task
environment.
Similar biases in ventral striatal activation are seen in other conditions such
as OCD and bias in processing losses (Jung et al. 2011), MDD and anhedo-
nia (Whitton et al. 2015), as well as SUD and cue reactivity (Volkow et al.
2010; Kühn and Gallinat 2011). Further, in a recent review of SUD, Moeller
and Paulus (2018) observed that ventral striatal activation patterns are related
to long-term abstinence outcomes: increased activation in response to drug
cues is related to worse clinical outcomes (i.e., increased substance use and
recidivism), whereas increased activation to monetary or nondrug reward cues
are related to better clinical outcomes. In each case, ventral striatal activa-
tion is biased in its sensitivity to a variety of environmental cues. Thus, the
dysregulated salience attribution does not appear to be a consistent, transdiag-

nostic insensitivity to reward or aversive stimuli. Instead, it appears there is an
inability to distinguish properly between task-relevant and task-irrelevant (or
intrusive) cues.
However, the question remains: How does such an increase in evoked ac-
tivity precipitate a perseverative or intrusive thought? One well-articulated
neurobiological mechanism, instantiated in the ventral striatum, for such a
transition to salience misattribution and uncontrolled attentional capture is
the glutamate-mediated transient synaptic potentiation model (reviewed by
Kalivas and Kalivas 2016).
Briefly, this model suggests that upon presentation of a previously potenti-
ated cue (e.g., a drug-related cue in SUD or an object of obsession in OCD),
excessive glutamine release in the nucleus accumbens (NAc) core leads to
transient synaptic potentiation that biases the attribution of salience to the po-
tentiated cue being processed. In addition, this bias in the NAc core leaves the
region less responsive to alternative cues that would normally be more fully
encoded (Kalivas and Kalivas 2016). This dual mechanism likely increases
the magnitude of the difference between potentiated and alternative cues and
further instantiates salience of the potentiated cue.
The molecular mechanism for such a transient biasing of NAc core processing
has been described in a rodent self-administration model: Drug-seeking behavior
was related to transient potentiation of D1 receptors in the NAc core following
presentation of drug cues. Moreover, following brief access to cocaine this poten-
tiation rapidly extinguished, only to be reinstated following 45 minutes of forced
abstinence (Spencer et al. 2017). These dynamics of potentiation are consistent
with models of SUD that describe preoccupation with and craving for drugs, both
of which can be viewed as NRTs (Koob and Volkow 2009).
Once a cue is ascribed a salience incommensurate with its task relevance,
transient synaptic potentiation within the ventral striatum may sustain this bias
by both increasing the coded salience of the potentiated (i.e., maladaptive)
cue, while simultaneously decreasing the salience of any alternative, compet-
ing (i.e., goal-directed) cue. Computationally, this cascade is consistent with
biased processing of specific subsets of previously encountered stimuli (e.g.,
drug cues in SUD), potentially leading to a reduced ability to modify posteriors
based on experience. Maladaptive ventral striatal activation across a variety of
psychiatric conditions is consistent with this interpretation.
Cortico-Striatal-Thalamo-Cortical Loops Convey

Salience Signals throughout the Brain
Regardless of how the NAc core encoding of salience is biased in disease,

neuroanatomical evidence clearly shows processing in the ventral striatum
does not occur in isolation. Dense and reciprocal interconnections between
striatum, thalamus, and cortex in the form of cortico-striatal-thalamo-cortical

(CSTC) loops allow for complex communication among brain areas when
processing salient and motivationally relevant information. These connec-
tions also provide an anatomical mechanism to convey reciprocally looping
ascending (striatum–cortex) and descending (cortex–striatum) influences
on salience computation. The physical connections along CSTC loops have
been described in great detail (Haber and Knutson 2009; Haber 2016), while
tract-tracing results in nonhuman primates can be clearly related to patterns
of resting-state functional connectivity in humans (Parkes et al. 2016; Choi
et al. 2017).
Classically, three loops linking the dorsal striatum with presupplementary
motor area, frontal eye fields, and dorsolateral prefrontal cortex (dlPFC), and
two loops linking the medial and ventral striatum with the OFC and ACC, re-
spectively, have been defined (Alexander et al. 1986). These connections form
a gradient in cortical projections to the striatum, whereby ventral striatal inputs
are associated with PFC areas processing emotion, caudate inputs with cogni-
tion, and putamen inputs with sensorimotor processing (Haber 2016). These
interconnections suggest that the biased signals processed in the ventral stria-
tum are ultimately conveyed throughout the brain, including specific regions
discussed below. Many of these are primary nodes of large-scale networks
across the cortex relevant to attentional control (Uddin 2014; Heilbronner and
Hayden 2016) and psychiatric disease (Menon 2011; Sutherland et al. 2012;
Kaiser et al. 2015).
However, it is important to note there is no clear anatomical or functional
boundary between the ventral and dorsal striatum, and the cortical projec-
tions to and from the striatum create a continuum of connectivity (Haber and
Knutson 2009; Choi et al. 2017; Marquand et al. 2017). Thus, assigning a
one-to-one connectivity relationship between striatal and cortical regions is not
possible. In fact, it has been estimated that the terminal fields of dACC, ventro-
medial PFC, and OFC cover almost 25% of the striatum, which is an overrep-
resentation as compared to the cortical volume of the brain (Haber et al. 2006).
This overrepresentation is especially germane to the current discussion due
to the role of OFC and dACC in reward and salience processing, respectively
(Bartra et al. 2013). Instead of a well-defined gradient of CSTC loop connec-
tions, clear zones of integration as well as convergence are observed across the
striatum (Haber and Behrens 2014; Choi et al. 2017; Marquand et al. 2017).
Choi et al. (2017) describe a homology between tract tracing in nonhu-
man primates and resting-state functional connectivity in humans, illustrat-
ing clear delineations in the pattern of physical and functional connection
between the ventral and dorsal striatum. This pattern of connectivity agrees
with previous tract-tracing evidence in nonhuman primates (Chikama et al.
1997). Specifically, more ventral striatum is strongly connected to the dorsal
anterior insula (aIns) (Chikama et al. 1997) and dACC (Kunishio and Haber
1994; Parkes et al. 2016). In contrast, the rostral dorsal caudate is a hub of
integration, with projections to and from a variety of brain regions associated

with attentional control, including caudal inferior parietal lobule (IPL), ventro-
lateral, dorsolateral, and dorsomedial PFC, dACC, and OFC (Choi et al. 2017;
Marquand et al. 2017).
The observed dichotomy in striatal projections to cortex is consistent with
circuits differentially impacted by chronic exposure to drugs. Our group has
identified a ventral striatal–dACC circuit whose resting-state functional con-
nectivity is reduced in chronic cocaine users and a dorsal striatal–dlPFC circuit
whose connectivity is increased in the same group as compared to healthy con-
trols (Hu et al. 2015). Further, the balance between the up- and downregula-
tion of these circuits was correlated with DSM-IV-TR compulsivity symptoms
in cocaine users. Similarly, in first-episode SZ, reductions in ventral striatal–
dACC resting-state functional connectivity are observed and have been cor-
related with reported symptom severity (Lin et al. 2017). These findings show
that ventral striatal–dACC connectivity modulations are relevant across mul-
tiple conditions associated with NRTs (in this example, SUD and SZ). Further,
the relationship between connectivity strength and clinically diagnostic criteria
suggests a role for these circuits as potential neuroimaging biomarkers of dis-
ease severity.
The CSTC loops between cortex and striatum are interspersed with connec-
tions from striatal regions to various thalamic nuclei, which in turn are con-
nected back to the cortex. The cortical and thalamic connections to the striatum
are coordinated, meaning interconnected cortical and thalamic regions both
project to the same striatal area. For the purposes of this discussion, the cen-
tral-medial (CM) and medial parafascicular (PF) nuclei of the thalamus, which
are connected to medial PFC areas including dACC (Behrens et al. 2003), are
also connected to the ventral striatum (Van der Werf et al. 2002).
An illustrative example of impairment across nodes of the described CSTC
loops in the processing of highly salient stimuli is seen in an imaging para-
digm that employs erotic pictures (Metzger et al. 2010). When the salience of
an anticipated erotic picture is processed by healthy participants, the nodes of
this CSTC loop linking dACC and ventral striatum via the CM/PF thalamus
become activated, as identified via high field fMRI. Here, CM/PF thalamus,
dACC, and aIns showed increased activation during the anticipation of a sa-
lient, erotic picture (the ventral striatum was outside of the field of view in this
study; Metzger 2010). Gola et al. (2016), provide further supportive evidence
by showing enhanced anticipatory processing of erotic pictures within the ven-
tral striatum of men seeking treatment for problematic porn use. Treatment
seekers in this later study displayed enhanced striatal activation to the cues
predicting erotic pictures but not to cues predicting monetary gains.
Taken together, Metzger et al. (2010) and Gola et al. (2016) show that an-
ticipation, not consumption, of highly salient stimuli increases activation in each
node of the CSTC loop (i.e., ventral striatum, CM/PF thalamus, and dACC).
The observed activation within these nodes during anticipatory processing is
consistent with the conceptualization of salience as a predictive computational

process for encoding anticipated relevance of a specific stimulus in the environ-
ment (Parr and Friston 2019), in this case an erotic image. These results provide
an additional example of biased salience processing in a behavioral addiction
(Potenza 2015; Kraus et al. 2016). This further broadens the scope of conditions
associated with NRTs and dysfunction in core nodes of these CSTC loops.
The role of the thalamus in health and disease is an area of active inquiry.
The traditional designation of the thalamus as a passive “relay station” for
information processed elsewhere in the brain is being reconsidered. Recent
evidence suggests thalamic influence on cortical connections, including the
coordination of activation and direct control of synchronicity between cortical
regions via gain control (Saalmann 2014), and active filtering of information
(for a review, see Halassa and Kastner 2017). Consistent with this more active
processing role, the mediodorsal thalamus is suggested to amplify signals in
the PFC (Parnaudeau et al. 2018) and to extend representations in the PFC
over longer time durations than those associated with cognitive processes,
like working memory (Pergola et al. 2018). Advances in thalamic parcellation
(Kumar et al. 2017) and quantification of its resting-state functional connectiv-
ity with the entire brain (e.g., via 7T fMRI) will only improve the resolution
of these findings and further articulate the role of thalamic nuclei in salience
attribution.
Prefrontal Cortex: Regions of Interest Implicated in Psychiatric Disease
With NAc core dysfunction now described along with neuroanatomical links
between this region and thalamic and prefrontal regions, including the dACC
and aIns, our focus shifts from striatal to cortical areas implicated in psy-
chiatric disease associated with NRTs. A recent meta-analysis of structural
(n > 15,000) (Goodkind et al. 2015) and functional (n > 5,000) (McTeague et
al. 2017) data from psychiatric patients across a variety of disorders associated
with intrusive thoughts (SZ, MDD, SUD, OCD) identified specific yet over-
lapping areas of dysfunction. Specifically, the pattern of gray matter loss in
patients was circumscribed to dACC and bilateral aIns (Goodkind et al. 2015).
In agreement with these structural findings, hypoactivation in cognitive control
task-evoked activity was also observed in the dACC and right aIns as well as
in the left dlPFC and right IPL (McTeague et al. 2017). The tasks employed
in the meta-analysis did not probe intrusive thoughts per se, but rather top-
down cognitive control more generally. That said, the pattern of hypoactivation
across disease conditions showed reductions in regions associated with both
attentional control (i.e, left dlPFC, right IPL) and the calculation of salience
(i.e., dACC, aIns).
The aIns and dACC are coactivated across a wide variety of cognitive con-
trol and attention-related tasks. In fact, activation in these nodes are among the
most observed results in the fMRI cognition literature (Behrens et al. 2013).
Given their ubiquity in the extant literature, aIns and dACC have been theo-
rized to play a central role in broad cognitive constructs, as in the detection
of relevant information from both the external and interoceptive worlds and
the coordination of appropriate attentional capture and behavioral response to
these salient signals (Uddin 2014; Heilbronner and Hayden 2016; Nour et al.
2018). The degree of task activation in these regions increases with the demand
for attentional control or with an increase in ambiguity of stimuli during per-
ceptual decision making (Lamichhane et al. 2016).
The insula is associated with the integration of interoceptive information
into calculations of salience (Critchley et al. 2004; Craig 2009) via ascending
pathways communicating visceral and allostatic information (Critchley and
Harrison 2013; Kleckner et al. 2017). Within the insula (and the dACC, as dis-
cussed below), these signals are integrated to create a single subjective image
of “our world” (Kurth et al. 2010); interoceptive representations in the insula
have been theorized to provide the basis for a perception of self via the inte-
gration of interoceptive signals related to physical state (Seth 2013; Namkung
et al. 2017). This self-focused processing localized to the insula is strikingly
consistent with the description of the content of repetitive thought as being
“related to one’s self and one’s world” (Segerstrom et al. 2003).
Returning to a transdiagnostic theme and the observation of insular hypo-
activity (McTeague et al. 2017), impairments in interoception are implicated
across psychiatric conditions associated with NRTs (Khalsa et al. 2018). Insular
activation in response to cued recall of a previously interoceptive challenge was
diminished in subjects with MDD as compared to healthy controls (DeVille et
al. 2018). Similar impairment in interoceptive processing focused on insular
hypoactivation (Naqvi and Bechara 2010) has also been associated with SUD
(Goldstein et al. 2009; Sutherland et al. 2013; Paulus and Stewart 2014).
In addition to a central role in the integration of interoceptive signals, the
insula is a primary adjudicator between external (exogenous) and internal (en-
dogenous) focus as a function of salience attribution (Sridharan et al. 2008;
Menon and Uddin 2010; Uddin 2014). Characterization of the causal inter-
actions between insula and dACC, however, shows additional divergence in
their patterns of activation. Specifically, aIns has been shown to amplify the
detection of salience within the dACC (Chen et al. 2014; Cai et al. 2015). This
timescale is consistent with EEG spectral analysis showing insular activation
precedes that of dACC (Chand and Dhamala 2016).
In contrast to the association with interoceptive processing in the insula,
the dACC is associated with cognitive monitoring and control processes along
with economic decision making (Botvinick 2007; Kolling et al. 2016; Shenhav
et al. 2016; Alexander and Brown 2017). Recent integrative accounts suggest
that the primary function of the dACC is to process multiple facets of informa-
tion about the context in which a decision is being made to enable the appro-
priate goal-directed strategy (Heilbronner and Hayden 2016; Li et al. 2018b).
That is, dACC codes task-state information (originating either endogenously
or exogenously) that is relevant to the current demands on the individual via

adaptive coding to discount or ignore irrelevant details (Heilbronner and
Hayden 2016). An important difference between these context-relevant cal-
culations and those occurring in the OFC, which are more closely linked with
reward, is that while the OFC encodes the value of the current choice, the
dACC integrates across multiple dimensions and across a longer timescale
(Kennerley et al. 2011).
The related calculations of interoceptive salience and context-relevant
stimuli within the aIns and dACC can be more formally combined via recent
studies of belief updating in healthy populations. The belief updating calcula-
tion incorporates the integration of new, relevant information with existing
expectations, termed Bayesian surprise (Barto et al. 2013). Divergence be-
tween prior belief and updated posterior beliefs characterizes the level of this
“surprise.” Recent work in healthy individuals shows that surprising (salient)
information relevant to updating beliefs is encoded in the ventral striatum,
aIns, and dACC whereas merely surprising but uninformative information is
not (Nour et al., 2018). This is an important distinction, as these regions ap-
pear to differentiate the information content of different sensory streams. This
function is in line with the construct of salience, as defined by Parr and Friston
(2019): the anticipated reduction in uncertainty is associated with a specific
element in the environment.
Thus, the increase in activation across the striatum, aIns, and dACC can be
putatively related to the accurate identification of relevant information (i.e.,
the salience definition described above) used to refine beliefs and guide goal-
directed behavior. Thus, along with the ventral striatum, the aIns and dACC
play a combined role in distinguishing novel (unexpected, uninformative) and
surprising (unexpected, informative) sources of information. Informative in-
formation is used to alter response strategies (e.g., attentional modulations or
motor output), whereas surprising but irrelevant information is not.
Applying this idea more directly to psychiatric conditions that are linked
to NRTs, a failure to accurately encode Bayesian surprise is consistent with
the hypoactivations in aIns and dACC described by McTeague et al. (2017).
Hypoactivation in these regions points to an inability to update beliefs effi-
ciently, which in turn may perpetuate a positive feedback loop whereby poor
estimates of salience guide attention to maladaptive elements in the environ-
ments, leading to perseverative focus on uninformative stimuli, which may
precipitate NRTs.
The Salience Network in Intrusive Thought
Within the described CSTC loop, dysfunction and/or maladaptive bias to-
ward potentiated stimuli is observed at multiple levels of salience processing.
Additionally, many of the implicated regions of interest are regularly activated
in concert depending on task demands. For example, Nour et al. (2018) show
ventral striatal, dACC, and aIns activation in the processing of informative,
novel information, with the dACC and aIns regularly coactivated (Behrens et
al. 2013). These commonalities suggest a benefit to examining the brain at a
higher level of organization, moving from regions of interest to dyadic circuits
to large-scale network organization.
To this end, dACC and aIns are the primary nodes of the salience network
(SN). Originally described by Seeley et al. (2007), and subsequently replicated
using a variety of methodologies (Dosenbach et al. 2007; Smith et al. 2009;
Power et al. 2011), the SN is a centralized processor that ascribes salience
to stimuli (Uddin 2014) while coordinating attentional resources between an
internal and external focus in response to task demands (Sridharan et al. 2008;
Menon and Uddin 2010). Such a function is clearly consistent with the roles
ascribed individually to aIns and dACC detailed above.
While the dACC and aIns form the primary nodes in the SN, additional
large-scale networks are broadly relevant to cognition (Smith et al. 2009; Ji et
al. 2019). The interaction between the SN and these networks leads to a more
global view of the dysfunction associated with intrusive thought; such SN in-
teractions can be conceptualized within a tripartite network model. Briefly, this
model describes a default mode network (DMN), which includes the rostral
and posterior cingulate cortices, parahippocampal gyrus, and bilateral infe-
rior parietal cortex (Raichle et al. 2001; Buckner and DiNicola 2019), and an
anticorrelated executive control network (ECN), which includes the bilateral
dlPFC and parietal cortices (Honey et al. 2007). During interoceptive process-
ing, the DMN is relatively more active and the ECN is relatively deactivated
(Fox et al. 2005; Keller et al. 2013). During exteroceptive processing the re-
verse is true. Toggling between these two networks is thought to be mediated
by the SN (Menon and Uddin 2010).
Indeed, regardless of the cognitive function ascribed to the individual re-
gions of the SN, transdiagnostic findings implicating nodes of the SN, both
structurally (Goodkind et al. 2015) and functionally (McTeague et al. 2017),
have led to conceptualizations of SN dysfunction as a core transdiagnostic
symptom of psychiatric dysfunction (Menon 2011). Transdiagnostic differ-
ences in activation within nodes of the CSTC loops connecting striatum to SN
nodes have been described in detail by Peters et al. (2016).
Tripartite Network Model and Aberrant Cognition
An influential theoretical model by Menon (2011) describes a transdiagnostic

hypothesis of aberrant salience calculation in psychopathology that centers on
dysfunctional network connections between SN, DMN, and ECN. The model
suggests that the normal adjudication between internal (DMN) and external
(ECN) focus mediated by the SN (Sridharan et al. 2008) is commonly disrupted
across psychiatric diseases. While entirely consistent with biased calculations

within the nodes of the SN (aIns and dACC) and ventral striatum described
above, this network-level model provides a systems neuroscience description
of brain dysfunction and explicitly relates changes in salience attribution to
emotional and attentional processing distributed throughout the brain.
The dynamic interactions between large-scale networks are determined by
a variety of factors. For example, the integrity of connections within a given
network (Honey and Sporns 2008; Boes et al. 2015) is important to determine
the functional capabilities of that network and the communication fidelity both
within and between networks. In addition, high-fidelity messages can fail to be
communicated successfully if hubs of interconnection between networks are
dysfunctional in disease (Cole et al. 2013; Gratton et al. 2018). These com-
munications also occur dynamically, and the engagement or disengagement of
functional connectivity between networks, a process mediated by the SN, are
potential points of failure at a systems level.
Menon’s model of dysconnectivity has proven prescient (Menon 2011).
Across a variety of diseases associated with NRTs, network-level dysconnec-
tivity between SN–DMN, SN–ECN, and DMN–ECN have been observed.
Using the tripartite network framework, a meta-analysis (n > 16,000 total pa-
tients, including n > 8,000 patients) across a variety of psychiatric diseases
(including MDD, OCD, PTSD, SZ) associated with NRTs identified transdi-
agnostic patterns of seed-based functional dysconnectivity both within nodes
of the SN, DMN, and ECN as well as between the three networks (Sha et al.
2019). Hypoconnectivity is observed at rest across diseases within nodes of the
DMN and SN and between nodes of SN and both DMN and ECN. In contrast,
hyperconnectivity is observed within the ECN and distinct nodes within DMN,
between the ECN and DMN, and between DMN and subunits of the SN.
These results are broadly consistent with Menon’s tripartite model: reduced
connectivity between SN and DMN or ECN suggests a reduced complement
of information to integrate into accurate and/or flexible estimates of prior and
posterior beliefs about the environment. However, hyperconnectivity between
DMN and nodes of the SN are also observed, suggesting the balance between
connections, as opposed to a unitary increase or decrease between networks,
may be a distinguishing feature (e.g., Hu et al. 2015).
A second recent result further broadens the scope of analysis: interrogat-
ing brain-wide connectivity patterns via a connectome-wide association study
approach (Shehzad et al. 2014) to identify brain regions whose whole-brain
connectivity pattern is associated with the p factor, a hypothesized common
factor underlying psychopathology across disease conditions (Caspi et al.
2013). Using a data-driven approach and a large data set (n > 600), four re-
gions in the occipital cortex were identified where whole-brain multivariate
connectivity patterns were correlated with p factor scores (Elliott et al. 2018).
These results are not directly interpretable within the tripartite network hy-
pothesis, as the regions identified fall outside of traditional SN, ECN, DMN
boundaries, when the identified occipital regions are used as seeds in a rest-
ing-state functional connectivity analysis, similar to those analyzed by Sha et
al. (2019). Nonetheless, hyperconnectivity with both the ECN and DMN was
positively correlated with p factor score. Importantly, while DMN and ECN
within the tripartite network model are usually considered oppositional (Fox
et al. 2005; Sridharan et al. 2008; Menon and Uddin 2010), the connectivity
of both networks was similarly enhanced with an independent, and psychiatri-
cally relevant, region of the occipital cortex. Further, the degree of this shared
increase in functional connectivity correlated positively with p factor score.
These results point to network interactions beyond the SN (Peters et al.
2016) that may indirectly influence the salience calculations instantiated within
dACC and aIns. In both of these recent cases, the network structure separat-
ing ECN and DMN appears to be reduced, either through hyperconnectivity
between these two normally oppositional networks (Sha et al. 2019), or via
increased coherence with a mediating node outside of the tripartite network
(Elliott et al. 2018). The coherence between ECN and DMN may further bias
the information integrated in SN, though an empirical demonstration of this
remains outstanding.
Conclusion
In psychiatric conditions that include NRTs as a symptom, a common set of bi-

ases in information processing and dysconnectivity between specific nodes as
well as large-scale networks is observed. In each case, these dysfunctions ap-
pear to reflect a failure to tune or modulate the response or connection properly,
as opposed to a broad deficit in either task-evoked activation or connectivity.
Recent advances in gathering large data sets across a diversity of psychiatric
conditions have aided in revealing these dysfunctions.
It is important to note that healthy cognition includes the regular experi-
ence of intrusive thoughts. It is an increase in the perseverative focus on these
thoughts that leads to clinically relevant dysfunction. Thus, it is not the pres-
ence of a response within or a connection between brain regions that is indica-
tive of a disease as much as it is the inability to discriminate properly among
alternatives or determine the most relevant information to guide decision mak-
ing. We suggest that processing biases at the level of the striatum, thalamus,
insula, and dACC indicate computational dysfunctions during the Bayesian
predictive coding of salience.
As a representative example, we extend the model by Kalivas and Kalivas
(2016) to incorporate a more explicit role for thalamus, dACC, and aIns along
with large-scale networks within the tripartite model of brain function. Biased
processing of potentiated stimuli, at the expense of alternative stimuli, within
the ventral striatum is strongly linked to glutamate-mediated transient synap-
tic potentiation (Spencer et al. 2017). These signals are conveyed along the
well-described CSTC loops (Choi et al. 2017) to the thalamus, where recent
evidence suggests their representations in PFC may be amplified or sustained
(Parnaudeau et al. 2018; Pergola et al. 2018), potentially increasing bias.
Hypoactivation of nodes within the SN (Peters et al. 2016; McTeague et
al. 2017) leads to suboptimal integration of interoceptive signals (Kleckner et
al. 2017), which may be biased due to the allostatic load of psychiatric dis-
ease more generally (McEwen and Gianaros 2011). In addition, the integra-
tion of endogenous and exogenous information to identify and fully process
contextually relevant information (Heilbronner and Hayden 2016) is likely
biased by the observed hypoactivity within nodes of the SN. These process-
ing biases within SN lead to an inability to identify relevant information in
a given context (Parr and Friston 2019) or to update beliefs to modify atten-
tional strategies or behavior more generally (Nour et al. 2018). Finally, the
integrative calculations of the SN may be further impacted by alterations in
large-scale network structure (Sha et al. 2019), which further bias salience
calculations by reducing the fidelity of information communicated with the
rest of the brain.
In summary, bias at each stage of salience attribution leads to an over-
representation of potentiated stimuli as well as to an insensitivity to counter-
factual evidence, which normally signals the need to alter behavior. A better
understanding of the calculations instantiated within each of these regions,
and more importantly, a more holistic, systems-based picture of their interac-
tions, is likely to identify novel therapeutic interventions that will allow us to
mitigate the unconstructive consequences of NRTs and to treat the underlying
dysfunction.
Open Questions
To guide future enquiry, we conclude by highlighting three problem areas that

await clarification through future research. First, are NRTs the cause of the
psychiatric diseases described or only a symptom? While the current concep-
tualization of failures in Bayesian predictive coding computations is consistent
with the neuroimaging evidence of dysfunction in these conditions, few direct
links have been described between these regions and the subjective experience
of NRTs in patients (for a notable exception in SZ, see Sterzer et al. 2018).
This computational framework, however, provides testable hypotheses to de-
termine how the estimation and updating of beliefs about the environment may
be causally linked to the experience of NRTs across conditions.
Second, what are the limits of the neurobiological framework centered
on CSTC loops? Any discussion of salience necessarily implicates the entire
brain. Which key nodes have not yet been accounted for in the current concep-
tualization (hippocampus, amygdala, dlPFC)? Especially in the estimation of
beliefs, the central role of memory processes is currently underspecified.
Finally, what potential treatments do these circuits suggest? Given the

multiple levels of systems that are impacted—from D1 receptors in ventral
striatum (Roberts-Wolfe et al. 2018) to large-scale brain networks (Sha et al.
2019)—are multipronged treatments, such as simultaneous pharmacotherapy
(Kalivas and Kalivas 2016) and transcranial magnetic stimulation (Peters et al.
2016) more likely to succeed?
Acknowledgments
This work was supported by the National Institute on Drug Abuse, Intramural Research
Program and Center for Tobacco Products (U.S. Food and Drug Administration) Grant
No. NDA13001-001-00000 (to EAS).
11
Brain Networks for

Cognitive Control
Four Unresolved Questions
David Badre
Abstract
The last decade has witnessed a marked shift of emphasis in cognitive neuroscience
away from simple localization of function and toward the organization, coding, and
dynamics of brain networks. This is surely a healthy evolution of our science, and the
study of cognitive control has benefited from this shift, as much as any domain. Howev-
er, the emphasis on brain-wide networks for cognitive control has reopened some older
debates, once thought resolved, while also introducing some new ones. This chapter
focuses on four questions viewed as unresolved and fundamental because one’s par-
ticular answer to them commits to some basic theoretical differences regarding cogni-
tive control function: Are there one, many, or any networks whose primary function is
best described as cognitive control? Are the networks supporting cognitive control in
the brain “hub-like” or “hierarchical” in their intrinsic and extrinsic organization? Are
the networks for cognitive control modulatory or transmissive in the pathway from
thought to action? Does controllability apply at the level of cognitive function or brain
state? Each question is defined and relevant background is presented that could inform
a resolution.
Introduction
A longstanding problem in cognitive science and neuroscience concerns how

the brain supports cognitive control. In broad terms, cognitive control refers to
the set of mechanisms needed to organize our thoughts or actions to achieve a
goal, particularly when the behaviors involved are not well learned or habitual
(Stuss and Benson 1987; Logan and Gordon 2001; Miller and Cohen 2001;
Badre and Nee 2018; Badre 2020). Cognitive control allows us to strategically
select responses appropriate to our circumstances, to adjust our behavior on
204 D. Badre
the fly, and to adapt to open-ended problems and novel situations. It allows
us to sustain goal-directed behavior over multiple timescales and to withhold
inappropriate responses, even when those responses are prepotent, habitual, or
stem from the prevailing urges of the moment. Cognitive control function lies
close to the heart of human intelligence and ingenuity. It is also vulnerable to
deficits across many, if not most, psychiatric and neurological disorders, be-
ing at the base of many of the behavioral problems arising in those conditions.
Thus, understanding the mechanisms by which the brain supports cognitive
control is a problem of fundamental importance.
Understanding cognitive control is of direct importance for intrusive think-
ing, the definition and scope of which is addressed in detail in other chapters
of this volume. Most definitions, however, require that intrusions are unwanted
and are unrelated to our goals or the task at hand. Thus, control mechanisms
are an important means by which we both avoid intrusive thoughts and man-
age their impact. It follows that understanding the brain systems that support
cognitive control function will have important implications for intrusive think-
ing, both in identifying its sources and seeking its potential remediation. In this
chapter I review the brain networks that support cognitive control as a general
background for more direct consideration of intrusive thinking.
As with most domains of cognitive neuroscience, the last decade of research
into cognitive control in the brain has witnessed a shift away from a paradigm
of functional localization toward one of functional networks. Among the most
robust and important observations to emerge from the overall network approach
has been that sets of brain areas tend to covary mostly with each other and not
with other areas (Power et al. 2011; Yeo et al. 2011; Buckner et al. 2013). Further,
the structure of this covariation is not entirely due to spatial proximity. Rather, af-
filiated areas can be distributed in each lobe of the brain, whereas other areas that
are spatially contiguous may not affiliate. These basic properties have allowed
for definition of brain networks or clusters of areas that covary with each other at
different scales (Power et al. 2011; Yeo et al. 2011).
Here, I focus on four big questions that are provoked when one takes a net-
work view of cognitive control seriously:
• Are there one, many, or any networks whose primary function is best
described as cognitive control?
• Are the networks supporting cognitive control in the brain “hub-like”
or “hierarchical” in their intrinsic and extrinsic organization?
• Are the networks for cognitive control modulatory or transmissive in
the pathway from thought to action?
• Does controllability apply at the level of cognitive function or brain state?
Obviously, this is not intended as an exhaustive list of questions about control
networks. Rather, these are the kinds of questions that I find myself asking
routinely, whether in my own work or in reading about others’. No one has de-
finitive answers, and so these questions also remain contentious or unresolved.
Brain Networks for Cognitive Control 205
My goal is not to provide answers in this essay, though I will express my own
view. Rather, I will define each question and present some relevant background
in the hope of provoking further discussion.
Are There One, Many, or Any Networks

Whose Function Is Cognitive Control?
One of the oldest questions in the study of cognitive control or executive func-
tion is whether there exists one or many executive controllers in the mind and
brain, or if there are executive controllers at all. The majority view has mostly
been that, while there exists cognitive control function, it is not simply one
thing. Rather, what we call executive function or cognitive control actually
refers to a variety of specific functions and capacities.
Two camps reject this basic view. First, there are those who contend that
there is one central system for cognitive control or executive function and that
little to no decisive evidence exists for strong dissociations among subtypes of
cognitive control functioning. The second camp argues that control is emergent
from network processing in the brain, but that no particular area or network
of areas is best characterized as primarily supporting “control.” Finally, even
among those who agree that cognitive control exists and has many facets, there
has been little agreement about the exact type and number of these facets.
This core debate has unfolded in almost every domain in which cognitive
control has been studied: from behavior to individual differences to neuropsy-
chology to neuroimaging. Currently, it is playing out again in network neu-
roscience. I will devote some more space to this first question than the other
questions as it also provides an opportunity to summarize some background on
the networks relevant to cognitive control.
The Multiple Demand System: One Network to Control Them All
One reason that the unitary hypothesis has been so hard to falsify conclusively
is that it is often the null hypothesis (Aron et al. 2015). It predicts that in any
setting in which one attempts to locate a difference based on a type of cogni-
tive control, there will be no difference. Thus, any imprecision in design, logic,
or measurement has the potential to find evidence consistent with this unitary
view by virtue of being inconsistent with the alternative. As a consequence,
the unitary view has been something of a “zombie hypothesis” over the years:
falsified in experiments that show dissociations in the brain or behavior, only
to rise again a few years later when the same distinction is not found to gen-
eralize to a new task or the methodology changes. However, it is important to
acknowledge that a failure to locate a difference, even in direct replication, is
not itself positive evidence for a unitary controller. Rather, unitary controllers
need positive predictions and evidence of their own.
206 D. Badre
In this light, the definition of the multiple demand system put forth by
John Duncan and colleagues is appealing as a unitary controller view of brain
organization, because it is based on a positive prediction: the multiple de-
mand system is engaged when you perform any challenging or difficult task
(Duncan 2013). Under these difficult circumstances, one needs to sequence
the set of attentional states required to perform the task. It is also in these
“hard” settings where one should expect the unitary cognitive control system
to be engaged. Importantly, however, the specifics of the task in question or
the demands that made the task difficult are not important. This system should
be fundamental and domain general, so that it participates across these differ-
ent task settings.
To test this hypothesis, Fedorenko et al. (2013) conducted an fMRI experi-
ment in which they contrasted difficult versus easy conditions in a wide range
of tasks. Difficulty was simply defined as a condition that took longer and
induced more errors behaviorally. The tasks differed in their specific demands
and the domain of input, such as between verbal or spatial. Nonetheless, when
one contrasted the hard with the easy conditions of these tasks, a consistent
set of areas was activated in each participant, as shown in Figure 11.1a. Given
its definition, this network was dubbed the “multiple demand system” or MD
system (Fedorenko et al. 2013).
The MD system has been studied extensively. It includes premotor cortex,
lateral prefrontal cortex (PFC) around the inferior frontal sulcus, the intrapa-
rietal sulcus, the anterior cingulate cortex (ACC), the frontal operculum, and
subregions of the basal ganglia (Fedorenko et al. 2013). This network has been
associated with a variety of measures of flexible behavior, including general
intelligence (Woolgar et al. 2010) and novel rule following (Tschentscher et al.
2017). In addition, most recently, it has been found to line up with the Human
Connectome Project parcellation that is defined based on a range of structural
and functional anatomical features (Assem et al. 2020).
As a unitary system, the MD system is proposed to serve a very gen-
eral control function needed across multiple complex tasks; namely, the
assembly of attentional episodes that are the smallest unit chunks of a
complex problem (Duncan 2013). When people seek to solve a new or
difficult task, it has long been thought that they must break the problem
into parts (Newell 1990). From the MD theory, each part is defined by a
set of input-output relations that are coordinated by attentional systems.
The MD system is proposed to manage these attentional episodes and the
transitions from one to the next. Thus, neural coding within this network is
thought to be highly dynamic, changing from moment to moment in a tra-
jectory determined by the flow of attentional episodes. The consistent and
widespread observation of flexible and dynamic coding of prefrontal neu-
rons from electrophysiological recording in the nonhuman primate shares a
qualitative correspondence to this view of multiple demand coding (Rainer
et al. 1998; Stokes et al. 2013).
(a)
Premotor cortex
Intraparietal Pre-SMA / anterior cingulate
sulcus
Inferior frontal
sulcus
Anterior insula /
frontal operculum
(b)
CO
FP
Figure 11.1 Networks activated across multiple task demands. (a) Activated regions
of multiple demand systems: contrast of hard versus easy conditions in all tasks run
(after Fedorenko et al. 2013). (b) Frontoparietal (FP) and cingulo-opercular (CO) net-
works defined through functional connectivity: different methods of network definition
find convergent network definitions (after Gratton et al. 2018).
Frontoparietal Control System and the Cingulo-Opercular,

Other Control System
As already noted, the functional definition of the MD system encompasses a

wide and consistent set of frontal, parietal, and subcortical regions. However,
evidence from analysis of functional connectivity in large resting-state data
sets indicates that these areas are separable into at least two networks: a
frontoparietal (FP) network and a cingulo-opercular (CO) network (Power
et al. 2011; Gratton et al. 2018). Whole brain parcellations repeatedly lo-
cate differences in connectivity between these two networks across multiple
methods, in large samples, and repeatedly in “deep sampled” fMRI subjects
(Power et al. 2011; Yeo et al. 2011; Gordon et al. 2017; Gratton et al. 2018;
Ji et al. 2019). Data from patients with brain damage to either regions of the
FP or CO networks exhibit reduced functional connectivity at rest within that
network but not across the networks, amounting to a double dissociation of
208 D. Badre
functionally connected networks (Nomura et al. 2010). Further, this network

distinction does not depend on studying connectivity at rest. A recent fMRI
study reproduced this network difference in connectivity within the function-
ally defined MD system while participants performed a cognitive control
task (Crittenden et al. 2016). Thus, the evidence is quite strong for a network
distinction in functional connectivity among two major networks that make
up the MD system.
Importantly, however, while the evidence for two networks within the MD
system suggests distinct functions are served by these two networks, the evi-
dence for what those functions might be is neither strong nor clear. A meta-
analysis of executive function tasks proposed a functional distinction between
the FP and CO networks based loosely on timescale of control (Dosenbach et
al. 2006, 2007, 2008). This analysis noted that the FP network was activated
in tasks involving task cueing or adjustments of a task from feedback. The CO
network, by contrast, was activated for these features in addition to demands
to sustain control over time. Based on these observations and follow-up work,
Dosenbach, Petersen, and colleagues proposed a distinction between “control
implementation” by the FP network and “task set maintenance” by the CO
system (see Gratton et al. 2018). These functional designations are intuitive,
but they are not specified in a concrete mechanistic or process-specific way.
To date, no study has cleanly operationalized these processes and pitted them
against each other. Thus, no evidence of a functional double dissociation be-
tween control implementation and task set maintenance presently exists for the
FP and CO networks.
It is notable in this context that other prominent frameworks have attributed
more mechanistic functional differences to the lateral PFC and dorsal ACC
areas that overlap with the FP and CO networks, respectively. For example,
Botvinick proposed that the dACC may be important for detecting conditions
that require control, such as response conflict, and thereby signaling upreg-
ulation of control signals by lateral PFC (Botvinick et al. 2004). Recently,
Shenhav et al. (2013) updated the conflict detection model to suggest that ACC
computes the expected value of control, a signal that specifies the type and in-
tensity of control carried out by lateral PFC. Others have proposed that dACC
has access to stimulus-response policies which allows it to make predictions
and detect errors in response outcomes (Alexander and Brown 2011). The pre-
dicted response-outcome model captures this mechanism and can account for
a wide range of results from both electrophysiology and neuroimaging. Still
other models have suggested that dACC plays a role in computing value of
counterfactual plans to be executed in the future (Fouragnan et al. 2019). Thus,
several models propose a functional distinction between dACC and lateral
PFC, which might extend to the FP and CO networks, though there is presently
little agreement about what these differences might be or consistent empirical
evidence for these distinctions.
Hierarchical Control and Distinctions within the Frontoparietal System
Within the broadly defined FP network there is evidence for further func-
tional distinctions and subnetworks (e.g., Dixon et al. 2018). These distinc-
tions have been most consistently observed in the context of complex tasks
that are designed to test hierarchical cognitive control (Badre and Nee 2018).
Hierarchical cognitive control refers specifically to cases where we must con-
trol actions based on immediate contextual signals, while also being influenced
by higher-order superordinate control signals that are either more abstract pol-
icy or extended in time.
In general, if a task requires tracking multiple contextual signals to keep over-
lapping behavioral policies separate, demands on hierarchical control grow. For
example, in a recent experiment, children and adults were instructed with a set of
mappings between cartoon characters and left or right button presses, the “Go”
task (Verbruggen et al. 2018). Prior to performing the Go task, however, partici-
pants were asked to view all of the cartoon characters, pressing the right button to
advance to the next character (the “Next” task). This meant that while perform-
ing the Next task, participants would occasionally press the rightward arrow to a
character that required a left response later on during the Go task. Such an over-
lap of responses can cause conflict, evidenced in slowed response time during the
Next task. However, this conflict, is reduced if one can impose a latent context
that separates the episode of the Next task from the later episode of the Go task
and their respective response sets.
Interestingly, when doing the Next task, children exhibited more conflict
than adults; children had a harder time imposing this context episode on the
task. Notably, this conflict was evident even though they had never performed
the Go task and were only instructed on the response rules for this task. So,
it was not rule following that was a problem for the children, perhaps counter
to the widely held view. The conflict indicates they immediately implemented
the rules just from the instruction. Rather, their slow response was a symptom
of diminished hierarchical control capacity: they could not keep the latent task
contexts separate.
Studies testing hierarchical control have consistently exposed differences
within the FP control system (Figure 11.2a). Across a range of studies us-
ing fMRI, transcranial magnetic stimulation (TMS), and testing of patients
with frontal lobe lesions, differences in policy abstraction (defined in terms
of the number of conditions or branches in a decision tree between stimulus
and response) yield differences along the caudal to rostral PFC, with the
highest levels of abstraction associating with the rostral mid-dorsolateral
PFC (Koechlin et al. 2003; Badre and D’Esposito 2007, 2009; Nee and
D’Esposito 2016, 2017; Badre and Nee 2018). Further, manipulations of
temporal abstraction, which refers to the degree to which a goal or task must
be held pending over time, have found fMRI activation in the most rostral
portion of the frontal cortex, the rostrolateral PFC (Koechlin and Hyafil
210 D. Badre
(a) Contextual (b)

Schematic control Sensory-motor
control control
(c)
First order Second order Third order Fourth order
0.5 2
Figure 11.2 Relationship of functional studies of hierarchical control and brain

networks defined from functional connectivity. (a) Results from a meta-analysis of
hierarchical control studies (after Badre and Nee 2018). The colors distinguish three
functional zones related to different hierarchical control demands related to using simple
(sensorimotor control) or complex (contextual control) contexts to control responses.
Schematic control refers to studies that manipulated temporal abstraction or subgoaling
and branching demands. Small shapes are individual studies. Large shapes are average
coordinates. Within the contextual control zone, spheres refer to second-order control
and diamonds to third-order control and show a further separation rostral to caudal of
these studies. (b) The 17-network parcellation from Yeo et al. (2011) with the three
networks most overlapping the three zones highlighted (after Badre and Nee 2018).
(c) The direct comparison of the Yeo et al. (2011) network parcellation with activation
across four levels of hierarchical control, from Badre and D’Esposito (2007), shows the
consistent network overlap in multiple lobes of the brain (after Choi et al. 2018).
2007; Desrochers et al. 2015; Nee and D’Esposito 2016; Badre and Nee
2018). Perhaps relatedly, the rostrolateral PFC has also been implicated in
tasks requiring information from memory, future directed thought, counter-
factual or alternative courses of action, or pending actions to act as control
signals (reviewed in Badre and Nee 2018). For this reason, Nee and Badre
gave this zone a general label of “schematic control” to emphasize its rela-
tionship with these types of computations.
Important to the present discussion, these distinctions along the lateral PFC
are mostly encompassed within the broadly defined FP network. However,
brain networks can be decomposed at multiple scales. Yeo et al. (2011) applied
a clustering procedure to a functional connectome collected at rest in a large
sample of participants. While one clustering solution, termed the 7-network
parcellation, agreed with the coarse FP- and CO-network distinction, they also
identified a finer-grained 17-network parcellation that broke up the FP net-
work into more than one network (Figure 11.2b). This included, for example, a
separate network for the rostrolateral PFC as distinct from the mid-dorsolateral
PFC and from the premotor cortex.
By comparing this 17-network structure with the functional delineations
identified by Badre and D’Esposito (2007) in the fMRI study of hierarchical
control (Figure 11.2c), Choi et al. (2018) found that there was a correspon-
dence between the functional bounds associated with task-based differences in
levels of hierarchical control and distinctions within the 17-network structure.
Further, there were also effects of hierarchical control in distinct regions of the
parietal cortex and medial frontal cortex in accord with the network structure
(Choi et al. 2018). In a direct comparison, it was found that network member-
ship, rather than rostrocaudal location, best predicted the hierarchical level of
a particular voxel (Badre and D’Esposito 2007). Thus, rather than a set of areas
or a gradient going from front to back along the lateral frontal cortex, ranked
by a factor like policy abstraction, Choi et al. (2018) found that there are a
set of subnetworks within the FP network (or MD network) that are differen-
tially activated, based on complex control demands such as policy or temporal
abstraction.
Frontostriatal Circuits and Gating Interactions
A further network property of control that has been highlighted by the study
of hierarchical control is the potential importance of corticostriatal loops
in controlling interactions between separate frontal circuits (O’Reilly and
Frank 2006; Collins and Frank 2014; Chatham and Badre 2015). It is well
established that the basal ganglia form a series of loops with the frontal
cortex via the thalamus (Alexander et al. 1986; Haber 2003). In motor con-
trol, these loops are thought to support a feedback-based gating function
(Mink 1996). Specifically, candidate actions represented by cell populations
in premotor cortex are initially too weak to fire, because thalamic drive is
under tonic inhibition by the globus pallidus. However, these candidate ac-
tions in premotor cortex also send descending inputs to the striatum. The
striatum, including putamen and caudate, receive broad inputs, not just from
this premotor region but from cortex more broadly. Cells in the striatum are
modulated by the presence of dopamine, which also induces plasticity so
that these cells can learn which combinations of actions and context have
been adaptive or not. Thus, the value of the actions considered in premotor
cortex is computed as a function of what is being processed in cortex more
broadly. If these actions have a history of being adaptive in this context,
212 D. Badre
“Go” cells in the striatum will elicit a cascade that ultimately disinhibits the
thalamus and allows the action to be output (Mink 1996; Wickens, 1993;
O’Reilly and Frank 2006).
One influential hypothesis is that these same corticostriatal feedback
loops can operate over goal and context representations that are needed for
cognitive control and that are maintained in working memory by the lateral
PFC (O’Reilly and Frank 2006). This computation is described using the
metaphor of a gate on working memory. When the gate is closed, informa-
tion does not pass in or out of working memory. When it is open, working
memory can be updated and top-down control signals deployed. The feed-
back loops of the basal ganglia could operate as these gates by controlling
transmission from one cortical network to another through their disinhibitory
action on the thalamus.
Consistent with this hypothesis, there is evidence from fMRI, patient, and phar-
macology studies for these corticostriatal interactions during tasks that specifically
manipulate input and output gating of working memory (Frank and O’Reilly 2006;
McNab and Klingberg 2008; Baier et al. 2010; Chatham et al. 2014; Chatham and
Badre 2015). Furthermore, the loops between the lateral PFC and the basal gan-
glia are ordered and topographic, such that there are both macro- and microlevel
loops between cortex and striatum that are arrayed in an orderly fashion along
the rostrocaudal dimension of the frontal lobes (Verstynen et al. 2012). Choi et
al. (2018) reported convergent evidence of hierarchical ordering within the stria-
tum in resting-state functional connectivity. Further, some evidence from fMRI
and TMS provides functional support for separate loops that control context- and
motor-level processing during rule learning and execution (Badre and Frank 2012;
Jeon et al. 2014; Korb et al. 2017).
Interaction among multiple corticostriatal loops is a candidate mechanism
for hierarchical control (see Figure 11.3; Frank and Badre 2012). Specifically,
the gated output of superordinate contexts maintained in working memory by
one network can act as a top-down influence on the corticostriatal gating loop
controlling subordinate networks. In this way, multiple contingent contexts can
interact hierarchically to control action.
Models of these multiple corticostriatal loop interactions have shown
that they can efficiently learn abstract hierarchical rules, transfer these struc-
tures to new tasks, and exhibit the same quasi-parallel decision dynamics
that humans employ when they perform hierarchical control tasks (Frank
and Badre 2012; Collins and Frank 2013; Ranti et al. 2015). Further, gat-
ing of contextual representations is a means of controlling input and output
through lateral PFC, thus breaking down hard problems into more manage-
able chunks. In this sense, these gating computations resemble Duncan’s
conception of an attentional episode (Duncan 2013). These computations
emerge, however, from an interaction among separate, hierarchically or-
dered subnetworks.
3rd-order loop 2nd-order loop 1st-order loop

Mid- Pre- PMd/
dlPFC PMd motor
Thalamus Thalamus Thalamus
GP GP GP
CN CN P
Figure 11.3 Schematic of a nested, interacting corticostriatal loop network for hierar-
chical control. The details of the corticobasal ganglia loops have been simplified in this
diagram. Each loop is a feedback loop for one cortical network. However, the output of
each network can act as a top-down influence on a lower-order loop. This nesting can
provide a mechanism for multiple-contingent gating needed for complex, hierarchical
control of behavior. Labeled areas are motor cortex (motor), dorsal premotor cortex
(PMd), anterior dorsal premotor cortex (pre-PMd), mid-dorsolateral prefrontal cortex
(mid-dlPFC), globus pallidus (GP), putamen (P), and caudate nucleus (CN). Reprinted
with permission from Badre and Nee (2018).
The Stopping Network
A rigorous and compelling line of research has associated a separate corti-

cobasal ganglia network with a distinct form of cognitive control from rule
following and hierarchical control, namely stopping. Inhibitory control has
long been a mainstay of cognitive control function. However, not all inhibitory
behavior (e.g., slowing, stopping, or withholding what we are doing) is the
consequence of an inhibitory process.
The distinction between inhibition as an outcome and inhibition as a coun-
termanding or stopping process has caused considerable confusion in the litera-
ture (Macleod et al. 2003). For instance, the Go/No-Go task commonly used to
study inhibitory control might tap into an inhibition mechanism that prevents
an urge to respond on No-Go trials. Not responding to a No-Go cue, however,
could simply reflect a decision not to go rather than an actual suppression of
a Go response. This ambiguity clearly poses a challenge to the study of the
systems underlying inhibitory control. Thus, to understand inhibitory control
in the brain, it is important to test cases where an inhibitory process is required
to stop an ongoing or initiated action or thought.
To test inhibitory control, the Stop-Signal task (SST) is the closest to a gold-
standard paradigm that we have (Logan and Cowan 1984). An action must be
selected on every trial of the task in response to a “Go” stimulus. However,
these initiated responses must be occasionally stopped when a “Stop” stimulus
onsets at a delay after the Go stimulus. Success on the SST will thus depend on
the deployment and intensity of an inhibitory process, measured behaviorally
as the Stop-Signal response time and correlated with individual differences in
inhibitory control, including relating to real-world impulsive behaviors such
214 D. Badre
drug addiction (Dalley and Robbins 2017). Not all impulsive behavior, how-
ever, is linked to inhibition tested by the SST.
Strong evidence from multiple sources has associated stopping in the SST
with a brain network that includes the right inferior frontal cortex, the pre-
supplementary motor area (preSMA), and the subthalamic nucleus (STN) (see
Figure 11.4; Aron et al. 2004, 2007; Aron and Poldrack 2006). These regions
are consistently activated in fMRI studies that employ the SST. Damage to the
right ventrolateral PFC and preSMA causes deficits in stopping that are dis-
sociable from other frontal regions, such as the dorsolateral PFC. Importantly,
the right ventrolateral PFC, preSMA, and STN interact as a dynamic network
to inhibit behavior (Aron et al. 2016; Wessel and Aron 2017). These regions
are connected by direct white matter connections, the integrity of which cor-
relates with the speed of stopping (Forstmann et al. 2012).
STN is a key node in this stopping network (Isoda and Hikosaka 2008; Li et
al. 2008; Schmidt et al. 2013). It projects an excitatory influence onto the glo-
bus pallidus, thereby enhancing its inhibitory influence over the thalamus. This
pathway can rapidly bypass the gating computations occurring in the cortico-
striatal loops and put the brakes on behavior. Recent evidence from an elegant
optogenetic study in the mouse confirms these basic features in the context of
the stopping that occurs during surprise (Fife et al. 2017). Specifically, excit-
atory stimulation of the STN cells that project to the globus pallidus caused
cessation of licking responses in a mouse. Then, inhibition of the STN elimi-
nated stopping due to a surprising stimulus.
The stopping network lies clearly distinct from the FP network involved in
contextual control that was discussed above (Aron et al. 2015). Even subcor-
tically, it appears most related to the distinct hyperdirect (rather than direct/
indirect) pathways through the basal ganglia. Thus, motor inhibition may be
another example of a dissociable form of control.
Further, there is growing evidence for a broader inhibitory role for this net-
work beyond countermanding motor actions. For example, we observed in-
creased theta band oscillations between preSMA and STN under conditions
of greater uncertainty, and this coupling correlated with slowing of responses
during the decision (Frank et al. 2015). Ostensibly through motor inhibition,
the impact of control was functionally at the level of decision making. By
stopping the output of a response, more evidence was allowed to accumulate
before committing to a response; this is formally equivalent to setting a higher
evidence threshold and making a more conservative decision. Finally, there
is evidence that components of the stopping network, including the right in-
ferior frontal cortex, may also inhibit cognitive actions, specifically the act of
retrieval from long-term memory (Guo et al. 2018; Castiglione et al. 2019). In
sum, there is a separate brain network for fast stopping, and there may also be
further subnetwork distinctions within this domain.
(a) (b) Premotor
M1 R-IFG
preSMA
ȕ
IFC STN
Motor
thalamus
GPi
Figure 11.4 Networks critical for stopping. (a) Cortically, the right inferior frontal
cortex (IFC), sometimes termed ventrolateral prefrontal cortex, and the presupplemen-
tary motor area (preSMA) have been consistently implicated as playing a causal role in
stopping during the Stop-Signal task (after Aron et al. 2007). (b) Schematic of the path-
ways between cortex, the subthalamic nucleus (STN), internal globus pallidum (GPi),
and thalamus that are thought to support fast stopping (after Aron et al. 2016). R-IFG:
right inferior frontal gyrus.
Control without Controllers
The evidence presented above supports either one or several networks involved
in control. However, a third perspective, most recently argued by Eisenreich et
al. (2017), holds that none of these networks truly supports cognitive control
as a unique function. Rather, since neurons are systems of distributed com-
putation, they have emergent features of control that arise naturally in such
systems. From this perspective, cognitive control is an emergent property of
network computation, and there is no specific system devoted to cognitive con-
trol in the brain (Eisenreich et al. 2017).
There are many examples of distributed systems in nature that display con-
trolled behavior without the presence of a central controller. Eisenreich et al.
(2017) gives the example of a bee swarm searching for a good site to build a
hive. Bees use dances to communicate to other bees that they have found, for
instance, a good hive site. More bees will come to the site and do the dance
if they agree with the location. Once enough bees are dancing at the site, the
dance changes to a “build here” decision. At that point, a decision threshold
has been passed, and the bees start to build. However, if there are multiple
sites, there is conflict. The bee swarms at each location grow more slowly, and
so more time is required to reach a decision. Importantly, this control adjust-
ment is carried out at a “swarm level,” not at the level of any individual bee, as
216 D. Badre
no bee is aware of both locations. Thus, distributed systems exhibit dynamics

that can be characterized as control.
To what degree is cognitive control similarly emergent? The strong version
of this perspective proposes that there is no population of neurons in the brain
that is devoted to representing a goal or that is directing actions toward it.
Rather, goal-directed behavior emerges naturally from the systems devoted to
action and perception and their local control dynamics. Control is distributed
throughout the brain rather than being a function with a locus, whether that
locus is a brain region or a network.
There is little positive evidence for the strong version of this viewpoint.
Rather, the primary evidence is negative, questioning evidence for cognitive
control networks in the brain and evidence of loss of function from brain lesions.
For example, Eisenreich et al. (2017) focus on the complex and nonlinear mixed
selectivity of neurons within the frontal and parietal association cortices in the
networks discussed above to question whether these represent a goal or context
as a top-down influence. In addition, Eisenreich et al. (2017) take a relatively
dim view of human fMRI and neuropsychology, noting that there are debates in
these literatures, or mixed results, regarding most proposed networks. They take
these debates to suggest that the evidence in favor of regions or networks that are
devoted to cognitive control function is inconclusive at best.
However, in my view, the evidence in favor of networks for control is not
as ambivalent as Eisenreich et al. suggest. The mere presence of complex and
nonlinear mixed selective cell coding constrains the mechanisms for control,
but it is not, in and of itself, at odds with these neural representations serv-
ing a control function. Likewise, Eisenreich et al. overstate the inconsistencies
found in the neuroimaging and neuropsychological literatures to some degree.
As described in the preceding sections, many findings are now quite consistent
and highly robust. The debates have boiled down to disagreements over func-
tional interpretations of fairly consistent distinctions. Nonetheless, from their
skeptical position, Eisenreich et al. make a crucial point not to underestimate
the inherent controllability of any distributed system. In theorizing, it is impor-
tant to distinguish this kind of distributed but local control from the centrally
organized, goal-directed control we associate with cognitive control function.
Are Networks Supporting Control “Hub-Like” or “Hierarchical”?
To some degree, merely recognizing that a network rather than a specific brain
area is important for a function like cognitive control commits the same shal-
low theoretical error as “blobology” did in the early days of cognitive neuro-
science. Labeling a network merely assigns it a location without providing
mechanistic insight or constraint on theory. However, the focus on networks
for control, rather than individual areas, does offer an opportunity to consider
new questions about the macro-level processing dynamics and functional
organization of those networks. Note that this is a distinct question from that
considered above, which is concerned simply with whether there are multiple
control-related networks and what their function might be. The question con-
sidered here is what the nature of the interaction among these networks, and
others, might be. In this respect, one question to emerge out of the study of
networks for cognitive control is whether these networks are hub-like or hier-
archical in organization.
Viewing cognitive control systems as hubs is an intuitive and appealing the-
oretical idea. In essence, cognitive control systems manage or modulate rout-
ing between other systems of perception and action to carry out tasks. Thus,
these networks are central to the network dynamics of the brain, will be active
across most tasks, and will exert broad influence. In other words, cognitive
control networks are flexible hubs, with near proximity to all other networks,
and with the ability to change their connectivity with multiple other systems as
needed to coordinate their dynamics during a task (Figure 11.5a).
Evidence from fMRI functional connectivity has provided some support
for the hub hypothesis. Cole et al. (2013) scanned participants while they per-
formed sixty-four different mini-tasks in the scanner. This procedure allowed
changes in connectivity to be assessed while people were shifting the rules and
domains over which they performed the tasks. Cole et al. observed that the FP
network showed the greatest variability in its connections with other networks
across all the tasks relative to any other network, including the CO network.
Furthermore, rather than just reflecting random variability in a small set of
connections, FP also had the highest participation coefficient, which derives
from how uniform its connections are across all networks. From these obser-
vations, Cole et al. (2013) concluded the FP network was acting as a flexible
hub, changing its connectivity based on the task and thereby modulating the
relevant network for a particular task. In subsequent work, this global cross-
task connectivity of the FP network has been associated with fluid intelligence,
a further clue to its potential importance for cognitive control, particularly dur-
ing rapid task instruction and execution (Cole et al. 2015).
To some degree, the flexible hub model resembles a unitary central con-
troller that is required to modulate all other dynamics in the brain. As already
noted, however, there are likely distinctions among networks for control, even
within the FP system itself. Indeed, a recent analysis of functional connectiv-
ity patterns of the FP network across multiple task conditions found that this
hub-like network was decomposable into at least two networks with different
patterns of connectivity, and that these patterns were similar to those identified
by Yeo et al.’s (2011) 17-network parcellation (Dixon et al. 2018).
An important alternative hypothesis to a global hub is that the subnetworks
for control relate to each other hierarchically, such that some networks ex-
ert higher-order influence over other subnetworks, which in turn exert control
over more restricted domains (Figure 11.5b). The cascade model (Koechlin
et al. 2003) essentially proposed such a dynamic along the rostrocaudal axis
218 D. Badre
(a) (c)
Sensorimotor
control
Contextual
control FEF
Schematic pre
PMd
control pre- Motor
motor
pre
Mid- PMv
(b) Hierarchical control dlPFC
IFJ
rlPFC
Figure 11.5 Hub versus hierarchical network organizations. (a) A hub network or-
ganization places a control network, like the frontoparietal network, at the center of
coordinating other networks where it serves a general and fundamental role in orga-
nizing all other networks. (b) A hierarchical network organization allows for multiple
controlling networks to share asymmetric influences with each other and to have dif-
ferences in their domains of control and proximity to other networks. (c) A schematic
summary of the results from Nee and D’Esposito (2016, 2017) showing hierarchical
interactions among frontal lobe networks (after Badre and Nee 2018). Regions along
lateral prefrontal cortex are shown within the three control zones referenced in Figure
11.2a. Heavy, unbroken arrows show strong directions of influence. Broken arrows
depict weak influences. Colored arrows are domain- or task-specific influences. Abbre-
viations: mid-dorsolateral prefrontal cortex (mid-dlPFC), rostrolateral prefrontal cortex
(rlPFC), ventral premotor cortex (prePMv), inferior frontal junction (IFJ) area, anterior
dorsal premotor cortex (pre-PMd), frontal eye field (FEF).
of the frontal lobe, such that abstract temporally extended control signals in
rostral frontal cortex influence more temporally proximate contextual signals
in lateral PFC, which in turn influence action control by premotor and motor
cortex. Other models of hierarchical control have shared similar dynamics,
including among nested corticostriatal loops and through medial PFC (Frank
and Badre 2012; Alexander and Brown 2015).
In a set of two fMRI experiments, Nee and D’Esposito (2016, 2017) pro-
vided evidence for a hierarchical structure within lateral PFC. These studies
used estimates of effective or directional connectivity from dynamic causal
modeling while subjects performed a set of complex tasks that engaged vary-
ing degrees of hierarchical control in verbal versus spatial input domains.
Hierarchical strength was defined in terms of greater outward than inward con-
nectivity (i.e., a region has broader outputs than inputs, as defined in Badre et
al. 2009).
The basic results from these experiments are summarized in Figure 11.5c:
Mid-dorsolateral PFC was active in higher- (more abstract) but not lower-
order tasks across both input domains. It exerted an influence on the more
caudal dorsal premotor cortex and ventral premotor cortex regions that were
active across both the simpler and more complex tasks, but only in the spatial
or verbal domain, respectively. These caudal contextual control regions also
received domain-specific input from sensorimotor regions. The mid-dorsolat-
eral PFC received greater input from the rostrolateral PFC during conditions
where temporal abstraction was required. In a follow-up TMS study, Nee and
D’Esposito replicated these findings and showed that stimulation of nodes in
this network produced behavioral effects that were broadly consistent with this
information flow.
These findings from fMRI in humans converge with earlier anatomical stud-
ies in the macaque monkey. Goulas et al. (2014) performed an extensive meta-
analysis of monkey anatomical projections using the CoCoMac database and
focused on the connectional asymmetry that might drive hierarchy. They coded
multiple sites in the PFC based on the same definition of hierarchy as above:
any area higher in the hierarchy would have broader efferent connections to
lower-order areas than the reverse. Consistent with Nee and D’Esposito, anterior
mid-dorsolateral PFC (areas 45 and 46) showed the greatest asymmetry on this
metric, relative to regions caudal to the mid-dorsolateral PFC or to the rostrolat-
eral PFC which is anterior to it (Goulas et al. 2014). Notably, although Goulas et
al. (2014) did find evidence that the mid-dorsolateral PFC was higher in terms
of this network definition of hierarchy, it was not the most hub-like, based on a
measure of betweenness centrality. This appears consistent with structural con-
nectivity metrics in humans as well (van den Heuvel and Sporns 2013).
It remains open how one should characterize the dynamics among networks
supporting cognitive control. The broadly defined FP system exhibits a hub-like
character, with high participation and flexibility in connectivity across multiple
tasks. There is also evidence that subnetworks within this overall system re-
late to each other hierarchically. In that system, there is no central domain
general hub. Rather, the rostral mid-dorsolateral PFC is not active or neces-
sary across all tasks; it is necessary during those complex tasks that require
higher-order contextual control. Lower-order areas within the FP system are
activated across more tasks, but they are domain specific. Thus, a hierarchical
control architecture assumes that global control of the whole system emerges
from limited, local, and hierarchical interactions among control networks. This
contrasts with a hub network that manages interactions broadly and globally.
Finally, it should be noted that there are other hierarchical models we
have not discussed that yield different organizations. For example, Barbas
and Rempel-Clower (1997) proposed a laminar definition of hierarchy which
220 D. Badre
distinguishes regions based on their output versus input layers of cortex. This
laminar definition of hierarchy also predicts hierarchical interactions within
the FP system, but places the more rostral areas, like rostrolateral PFC, higher
(Goulas et al. 2014). Thus, the architecture and organization of networks for
control remains a mostly open question at present, but at least one core distinc-
tion is between those proposing hub-based interactions and those proposing
hierarchical ones.
Are the Networks for Control Modulatory or Transmissive?
A core assumption in most brain and network theories of cognitive control is

that their function is modulatory rather than transmissive. This dichotomy was
highlighted by Miller and Cohen (2001) in their seminal review on the PFC and
cognitive control. Their claim was that the PFC does not lie along the pathway
from stimulus to action. Rather, a series of pathways from input to output exist
in the brain that differ in their various strengths of connection. Collectively, these
pathways represent the full action repertoire of the system. What PFC contributes
is a system set apart for maintaining contexts in working memory and deploy-
ing them as control signals that can bias competition among these pathways for
behavior. From this perspective, then, PFC is modulatory, not transmissive. As
such, one could remove the PFC and this would not prohibit actions from oc-
curring in response to inputs. However, as PFC maintains high-level goals and
contexts, its loss would prevent the system from selecting action pathways based
on abstract, temporally remote, or task set information that is not available in
the immediate stimulus. The result is primarily automatic behavior based on the
strongest stimulus-to-response mapping.
As already noted, most current perspectives on the brain networks for con-
trol take this modulatory view. These networks serve control by maintaining
contexts; then through gated hierarchical interactions or flexible hubs, they
bias the right organization of the system to carry out the task that will achieve
the desired goal or fits with that context.
There is renewed reason, however, to reconsider this accepted view, or at
least the strong version that the PFC selectively maintains a context representa-
tion required for modulating other systems that route inputs to outputs. To see
why, consider the problem of control as a route driven between two locations in
a town. A good control system is set up such that any start point can reach any
end point. This is often done by building some main roads through town that
everybody uses. This is a generalizable system because the right combination
of these roads can assemble any route. But, it also causes problems. As they
are general and everyone uses them, such roads are susceptible to traffic. Thus,
we have to add gates (traffic lights) and monitor where we are going. This is
analogous to the interference or competition among stimulus-response (SR)
pathways that we experience in a task that overlaps with other tasks because
we are using general rules. Now consider that you have a particular route that
is being used a lot. You could build an express road between those two loca-
tions in town: at some cost of time and asphalt, you then add some dimensions
to your road system and gain a low traffic route. Increasing your dimensional-
ity is costly but highly efficient, if you know you need a particular set of routes.
From this analogy, one way to think about the transition from controlled to
automatic behavior is to view it as a transition from a reliance on generaliz-
able, low-dimensional neural representations that are subject to interference
to high-dimensional neural representations that take time to build but which
directly map a combination of inputs to a response. These transformations
could occur as transformations within frontal systems themselves. Early on,
coordination among more networks, using gates and so forth, is necessary
because a new task has to be assembled from low-dimensional components.
However, over time, it is efficiently supported by a high-dimensional repre-
sentation that allows a more direct route from input to output. It is still routed
through the PFC, where multiple contexts and goals can affect it, but just
differently in terms of the format of the routing (e.g., from low to high dimen-
sions). This is different from the modulatory view which requires that there
are always the same separate tracks from input to output: control acts like a
switch operator deciding which track gets to run and when. This is among
the distinctions that Eisenreich et al. (2017) made in their argument about
emergent control systems and is captured by their schematic representations
of different control architectures (Figure 11.6).
The evidence for this transmissive rather than modulatory model of control
is limited at present, but there are intriguing clues. First, the computational
trade-off described above between generalizable low-dimensional representa-
tions versus parallel high-dimensional representations has been shown in theo-
retical work using neural networks (Fusi et al. 2016; Musslick et al. 2017).
Second, there is evidence from physiology in the nonhuman primate and mul-
tivoxel pattern analyses of human fMRI data that the FP network does not
encode single contextual features of tasks but rather large conjunctions of mul-
tiple task features (Woolgar et al. 2011, 2016; Rigotti et al. 2013; Pischedda et
al. 2017). Presently, we lack clear evidence that separate areas or networks rep-
resent separate contexts or elements of a task. Third, maintenance in working
memory may not be a fixed-point system, wherein information is maintained
in a single stable form to be accessed at any point as an external control sig-
nal. Rather, evidence from electrophysiology in monkeys and EEG in humans
suggests that neural ensembles undergo dynamic change over time (Stokes et
al. 2013). Thus, these representations are themselves expressed in trajectories
toward an end point. Finally, evidence from nonhuman primates has shown
that the nonlinear mixed selectivity of PFC neural representations supports
high-dimensional capacity during task performance (Rigotti et al. 2013). This
is what allows these populations to encode multiple mixtures of their inputs
in unique patterns that can be read out by downstream cells. Rigotti et al. also
222 D. Badre
Modulatory control
Controller
network
Processor
Stimulus A Action A
Stimulus B Action B
Transmissive control
Controller/processor
network
Stimulus A Action A
Stimulus B Action B
Figure 11.6 Schematic of the difference between modulatory versus transmissive

control networks (after Eisenreich et al. 2017). In the modulatory control network (top
panel), the contextual controller lies outside the pathways from stimulus to response.
Its influence is like a switch operator, choosing which path from stimulus to response
is enacted. Its removal removes controlled behavior, leaving behind only automatic
behavior. In the transmissive control network (bottom), control networks are a part of
the pathway from stimulus to response. The nature of these representations, however,
changes over the course of experience with a task. Thus, transitions from controlled to
automatic behavior are supported by features in the geometry of the population coding,
such as high- versus low-dimensional coding.
provided evidence that this high-dimensional coding is behaviorally relevant,

as trials in which a monkey committed an error were associated with a reduc-
tion in dimensionality.
It is important to emphasize that these results could be interpreted in several
ways. It could be that high-dimensional representations are why the PFC can
be flexible. In other words, high dimensionality allows multiple input states to
be mapped to multiple output states. Alternatively, as monkeys in these experi-
ments have been extensively trained on these tasks, they have automated the
task and formed direct mappings from multiple input states to output states in
a compact way. Regardless, both perspectives are largely transmissive in their
view of the PFC rather than modulatory. PFC is part of the routing, but the
nature of the routing is constrained by the format of its representation, either

high or low dimensional. These observations reopen the question of whether
control networks are best characterized as modulatory, sitting outside the basic
flow of perception to action, versus being a direct part of it but changing their
format and coding as a function of automaticity.
Does Controllability Apply at the Level of

Cognitive Function or Brain State?
The conception of the mind and brain as a control system is one of executive
function’s most animating theoretical ideas, dating back at least to Norbert
Wiener’s mental-servo notions in Cybernetics (Wiener 1948) and the semi-
nal studies on the human operator in motor control by Kenneth Craik (1948).
These ideas and their descendants rely on the engineering formalism of opti-
mal control theory. In a control system, there is a set point, which is the desired
system state, and mechanisms of feedback or prediction that lead the system to
adjust toward the set point either in response to or in anticipation of changes
to the system state.
In the classic example of a thermostat, the state of the system that mat-
ters is the temperature in your home. The set point is the desired temperature.
Feedback to the system in the form of temperature measurements can result in
heating or cooling actions that will change the temperature of the environment
until the set point is reached. This is feedback control. Fancier modern ther-
mostat systems may also anticipate or learn about how ambient temperatures
change over the course of a day. Such a system can engage proactive cooling or
heating to maintain a stable set point, and so implement feedforward control.
Regardless of its specifics, however, the efficiency with which a control system
can reach its set point and the range of set points it can reach are a means of as-
sessing its quality. Control systems can be evaluated, compared, and optimized
on the basis of their ability to reach any particular desired state from any initial
state (termed controllability) and the efficiency with which they do so.
In cognitive control, control theory concepts have been historically posed
at the cognitive-functional level. The set point is defined with reference to
some real-world defined goal or target, such as drinking coffee or making it to
your connecting flight or naming the ink color in a Stroop task. An effective
control system is one that allows us to reach the widest range of such goal
states efficiently, either in the world or our cognitive system, given a similarly
wide range of initial contexts and situations. In this conception, maximal con-
trollability (i.e., being able to get to any output state given any input state) is
presumably what cognitive neuroscientists intuitively mean when they use the
term flexible behavior.
From a control theory perspective, psychological or neural mechanisms
must gather feedback or make predictions about the distance to desired set
224 D. Badre
points at this cognitive-functional level. Then, some mechanisms or pro-

cesses are proposed that select and implement mental or physical actions
to reach them (e.g., moving, remembering, thinking, naming, inhibiting).
Learning is similarly based on feedback from the world about one’s current
functional-level state and the actions that were taken. In elaborating these
models, neuroscience focuses on the neural mechanisms that implement
these functional-level control operations.
The expected value of control model introduced above is an example of a
theory of cognitive control that emphasizes optimal control theory (Shenhav
et al. 2013). From this perspective, a control problem occurs when there is a
disparity between a goal state and the current state, such as response conflict
arising during a Stroop task. The control system, in this case the dACC, is
able to compute not only this distance but the mental effort needed to resolve
it, in terms of the type and intensity of the control signal needed. This is
the expected value of control; namely, the value of achieving the goal dis-
counted by the effort required to reach it. Decisions about what and when to
engage control, then, can be made optimally by the brain as a function of these
computations.
As with expected value of control, most models of cognitive control orga-
nize the control problem at this functional level and then draw links to neu-
ral mechanisms at different degrees of specificity. Recently, however, a new
set of perspectives on control have emerged within a network connectivity
framework and these emphasize a subtly different level of controllability. For
example, a line of sophisticated work has applied advanced network analysis
techniques of white matter connections to characterize the controllability of
the brain (Laurent et al. 2015a; Betzel et al. 2016; Gu et al. 2017; Khambhati
et al. 2018). In brief, these analyses have emphasized how the density and or-
ganization of connections affect transitions within the space of possible brain
states. Analyses within this paradigm have emphasized two kinds of transitions
(Laurent et al. 2015a): transitions to common, “easy to reach” states are associ-
ated with densely connected networks, like the default mode network, whereas
transitions toward rare, “hard to reach” states are facilitated by networks with
weak connections, such as the FP and CO networks (Figure 11.7). This sug-
gests that these control networks are well positioned for maximal controllabil-
ity, shunting the system into any desired brain state.
These exciting new ideas offer a powerful approach to understanding control
in brain networks, and the search for translational, developmental, and clini-
cal correlates of these metrics is ongoing (Cornblath et al. 2019). However, as
described, the underlying model of control in these cases differs fundamentally
from the traditional functional-level control systems described above. These
systems define the control problem at the level of brain state rather than cog-
nitive function or real-world goal state. In other words, the control problem
is not how do you get that particular drink you want, but rather how do you
get the brain into a particular state that corresponds to having that drink. The
0 234 0 234
Average controllability Modal controllability
1% 3% 3%
12% 2% 16%
11% 17%
11%
19%
15%
30% 7%
4% 1%
12%
18% 18%
Default mode Ventral attention Auditory

Frontoparietal Dorsal attention Visual
Cingulo-opercular Other Somatosensory
Figure 11.7 Analysis of network controllability in the structural connectome (after

Gu et al. 2015). Left: Networks show the highest average controllability, which reflects
how efficiently they can move to “easy to reach” brain states. The default network is the
highest on this metric. Right: Networks show the highest modal controllability, which
reflects the efficiency with which they can move into “hard to reach” states. Highlighted
in this analysis are the frontoparietal and cingulo-opercular networks.
set point, then, is a target pattern of brain activity, not a real-world objective,
and the control system must plot the distance between your starting pattern of
brain activity and that goal pattern of brain activity through a functional con-
nectome. Finally, controllability is not defined in terms of how you behave but
rather how readily you can shift from the brain state you are in to any desired
brain state.
So, why aren’t these levels of theorizing about control the same? Isn’t this
network conception just a reductionist reframing of the original functional-
level control problem in terms of brain states? This is certainly the way it is
often posed and interpreted. However, a complication arises because of the
classic philosophical problem of multiple realizability. This is exactly the case
226 D. Badre
where this abstract philosopher’s thought problem has some real implications
for scientific theory.
Multiple realizability was famously raised by Hilary Putnam as an argu-
ment against identity theories in which each psychological state had one and
only one implementation (Putnam 1967; Block and Fodor 1972). The argu-
ment is that multiple species or individuals within a species can realize the
same psychological process, like pain or vision, with very different brains.
Thus, there is a many-to-one mapping of brain states to psychological states.
David Marr’s idea of distinct levels of analysis relies on similar arguments
about the asymmetry as one goes from computational to algorithmic to imple-
mentational levels (Marr 1982). Essentially, the functional or psychological
level is abstraction over multiple possible realizations of algorithms and imple-
mentations within the brain.
The implications of this concept for cognitive control are important. For the
reductionist reframing to correspond directly to the functional-level models,
there must be a one-to-one correspondence between any one functional-level
state (e.g., drinking) and a corresponding brain state (or a highly correlated
class of such states). However, as the problem of multiple realizability high-
lights, this assumption is hardly guaranteed. Rather, the goal of having that
drink may entail a wide range of activities to get there and a wide set of pos-
sible realizations of actually quenching one’s thirst. Each of these is associated
with a set of brain states. Some may not be more strongly correlated with
each other than with other states, and so comprise a disjunct set. Thus, what
ultimately connects this disjunct class of brain states is the functional-level
outcome, drinking. As such, conducting control at the functional level would
be the best way to ensure success, rather than making specific brain states a
set point.
A further issue concerns feedback in a brain state control system. To work
at the brain state level, the control system needs a means of detecting its dis-
tance from its set point. But, we don’t have explicit access through the senses
into our actual brain state, in terms of what neurons are firing and when. We
do, however, have access to a functional-level description, like whether we are
drinking or not. This feedback is also essential for the control system to learn
and know what actions to take to reach a goal in the future. Without assuming
a direct correspondence between the functional and brain state level that vio-
lates multiple realizability, feedback about the specific brain state target is not
available to the control system in an obvious way. This is a problem if one’s
control system is operating primarily at a brain state level. We do, however,
have perceptual systems that can assess the real-world outcome of our actions
and can use these to assess our state.
What this discussion highlights is that computing the distance to a goal
and seeking inputs that minimize that distance is different, depending on
whether one is mapping the distance to a particular brain state or the distance
to a disjoint set of such states defined by a functional outcome. Even if one
allows for some correlation among that set of states, it is easy to see how the
problem gets quite complicated in the latter case, if the control system has no
access to the functional-level description and is instead optimizing control
over brain states.
These complications notwithstanding, the issue of level of controllability
remains unresolved. It is not clear to me that the conventional functional view
is necessarily always the correct one or whether both options might not be
true and influence matters under different circumstances. At the level of plan-
ning, awareness, and explicit control, the functional-level description of con-
trol might be the appropriate level at which to understand cognitive control
for the reasons discussed above. As such, learning and feedback mechanisms
must ultimately reference this level of analysis. Neural accounts must explain
how the brain supports this functional-level control system, deploying neural
mechanisms that interpret the state of the world with respect to goals, compute
distances to real-world hypothetical and counterfactual outcomes, define the
means to cross that space, and monitor progress as it goes. However, for other
kinds of control, such as switching among well-learned tasks or adjusting on
the fly to maintain a stable trajectory of behavior, things may be different. In
these cases, the principles and constraints of brain-level network control may
be the most relevant feature determining individual variability in success, even
for functional-level outcomes.
In sum, whether control plays out at a primarily functional versus brain
level is an important open question, particularly as we take more sophisticated
approaches to understanding dynamics within the brain’s connectome. These
levels of controllability are not mutually exclusive, as both may influence con-
trolled behavior. This discussion highlights the importance of being explicit
about the level at which we assume control is occurring in our theorizing,
and that it is not trivial to assume that brain state control is isomorphic with
functional-level control.
Concluding Thoughts
The emergence of network neuroscience has brought with it an opportunity

to reevaluate some older questions about control and to raise some new and
exciting ones. Answers to these questions frame most of the major theories of
cognitive control in the brain, with each of them staking out a position implic-
itly or explicitly along these lines. As I noted at the beginning, however, this
is hardly an exhaustive list of the major questions facing control theorists. For
example, neural dynamics clearly constitute a very important aspect of brain
processing, and how dynamics among networks relate to control is only start-
ing to be understood. One can easily think of other such questions. This essay
is merely a starting point for considering the implications that network neuro-
science holds for our understanding of cognitive control function and the level
228 D. Badre
of neural and psychological mechanism. With relevance to this volume, these

concerns will also constrain any conception of control of intrusive thought as
well as the ways one might intervene.
Acknowledgments
This work was supported by the National Institute of Mental Health (MH111737) and
National Institute of Neurological Disease and Stroke (NS108380) of the NIH, a MURI
award from the Office of Naval Research (N00014-16-1-2832), and an award from the
James S. McDonnell Foundation. I am grateful to Apoorva Bhandari, Olga Lositsky,
and other Badre Lab members for helpful comments and discussions on these topics,
and to Emily Chicklis for her help with figures and illustrations. Figure 11.7, from Gu
et al. (2015), and Figure 11.4b, from Aron et al. (2016), are used under the terms of the
Creative Commons Attribution 4.0 International License (CC-BY). Copyright for Fig-
ure 11.4a, from Aron et al. (2007), is held by the Society for Neuroscience
12
A Framework for
Understanding Agency
Kayuet Liu and Hakwan Lau
Abstract
Why do some thoughts feel involuntary and intrusive? When should we hold someone
responsible for their actions and thoughts when they all have some basis in the brain? Are
we truly free agents when we are bounded by shared values and culture? This chapter
presents a framework for how our consciousness of our own intentions and emotions
allows us to form causal narratives about ourselves and the world. These narratives de-
termine our sense of agency, and we ascribe responsibility correctly depending on the ex-
tent to which one is capable of forming culturally appropriate narratives. Different ways
of characterizing consciousness are analyzed, with a focus on one that may prove most
useful within the context of understanding individual agency. A variant of the higher-
order view of consciousness is advocated that allows us to form causal, albeit imperfect,
narratives about ourselves. However, it is because of these imperfect narratives that our
understanding of agency and responsibility is formed. Thus, understanding how these
narratives come about is an important first step to understanding agency and how some
thoughts are considered involuntary and intrusive. Implications of this framework are
discussed using examples from mental illnesses, addiction, suicide, and racism.
Do Our Brains Make Us Do It?
Aberrant acts committed by patients with severe mental illnesses (e.g., schizo-
phrenia) are often considered not punishable. By law, if patients have lost
their capacity to reason, society should not hold them criminally responsible
(Mobbs et al. 2007). However, there seems to be a spectrum of controllability
(Moscarello and Hartley 2017) within which we ascribe degrees of respon-
sibility to patients suffering from different mental disorders. Take addiction
as a contrasting example. Law aside, members of society often disagree on
whether addicts are responsible for their own actions. Some hold that it is the
addicts’ own “decision” to go down the path of becoming who they are. Some
have challenged the notion that addiction is a brain disease, because the neural
correlates of addiction are not sufficient to cause addicts to do what they do in
230 K. Liu and H. Lau
many environments (Levy 2013). The question, then, is: How “voluntary” are
the behaviors of addicts (e.g., seeking out a drug) compared to those of patients
with schizophrenia (e.g., talking to an imaginary friend in public), since both
have bases in the brain? Indeed, when a healthy student is late for class due
to procrastination rather than a traffic jam, this behavior is also based in the
brain. In a sense, people’s brains are the causes of their behaviors in all these
instances. Yet we ascribe agency and responsibility differently in each case.
Perhaps this relates to how we see an individual’s agency and responsibility
in the context of culture. If someone is brought up in a society in which it is
acceptable to take food off each other’s plates without first asking, much as we
may disapprove of such actions, we may accept that person’s behavior more
easily than had that person been brought up in a culture where such behavior
was sternly forbidden. In this sense, are we truly free individuals, acting volun-
tarily out of our own desires or judgments? Or are we bounded very much by
our shared values, so that our errors may reflect more on the failure of society
rather than ourselves?
In this chapter, we do not attempt to solve these difficult questions, but
rather hope to provide a useful framework within which they can be addressed.
We will argue that the notion of consciousness is crucial to understanding these
issues. However, there are many different ways to characterize consciousness.
We evaluate a few accounts and focus on one that may prove useful within the
context of understanding individual agency.
Classical Literature on Free Will
Traditionally, debates on free will concern whether our actions are predeter-
mined; that is, whether our actions are genuinely de novo. The assumption is
that, in principle, if we knew all current physical events of the world, together
with a complete understanding of all physical laws, we should be able to pre-
dict the next events perfectly (Laplace 1951; Hoefer 2016). Within this con-
text of determinism (Hoefer 2016), how can one be an “unmoved first mover”
(Strawson 1994; Pereboom 2001)? Are our actions not already fully determined
physically, before they actually take place? To the extent that some notion of
freedom is possible within the deterministic framework (i.e., compatibilism;
Fischer 2006; Nahmias 2016), it cannot be because these actions are random
and thus unpredictable. Rather, we are responsible for our actions because we
have some degree of autonomous control over our actions. To argue that our
actions are genuinely free in this sense, one option is to argue that the physical
world is not truly fully deterministic. Some have appealed to findings from
quantum physics: that certain future events are not fully determined, even if all
current physical events are known (Kane 1996). This debate is important and
interesting, but many good reviews of the literature already exist (e.g., Sinnott-
Armstrong 2008). Here we focus specifically on considering which cognitive
A Framework for Understanding Agency 231
architectures may allow some meaningful notion of control to happen, while

assuming that classical Newtonian physics is sufficient for understanding cog-
nitive systems such as ourselves. For an in-depth discussion of these issues,
with regard to the implications of quantum physics, see Tse (2013).
The Consciousness Requirement
Shifting from the traditional focus on whether our actions are predetermined,
it has been argued that for individuals to be held responsible for their actions,
they need to be conscious of certain relevant events (Caruso 2012; Levy 2014).
Intuitively, this makes sense: it seems unfair to hold one fully accountable for
something that one isn’t even aware of having done, nor having even remotely
contemplated doing. Incidentally, there is experimental evidence that this
“consciousness requirement” is in line with our folk psychological concepts of
free will and responsibility (Shepherd 2017). Accordingly, many have focused
on the question: To what extent do conscious mental states truly cause behavior
(Pockett et al. 2006)? This diverges from the traditional question of determin-
ism, because this new question is still meaningful even if consciousness itself
were fully deterministic (Nahmias 2014). So long as the (deterministic) con-
scious processes in the brain causally influence our behaviors, there may still
be an important sense in which we have control over our actions.
There are, of course, critics of the view that consciousness is relevant. For
example, Smith (2005) claims that forgetting a close friend’s birthday (i.e.,
something that one does not consciously choose to do) does not eliminate the
responsibility of failing to call or send a card. We will not go into the details
here (for further discussion, see Caruso 2012), but an important lesson to de-
rive from these exchanges is that the arguments often depend on which specific
notion of consciousness is at play. We will focus on this issue in the next few
sections.
The Surprising Power of the Unconscious
The question of whether the conscious processes in the brain are causal to be-
haviors may seem trivial (Baumeister et al. 2011). Although we all feel that our
conscious thoughts and decisions have causal efficacy, several lines of empiri-
cal studies seem to challenge this intuition. First, studies of unconscious prim-
ing, mostly coming from the area of social cognition, show that our actions and
decisions may be influenced by unconscious cues (e.g., words or symbols ir-
relevant to the primary tasks at hand, or the gender or age of the experimenter),
the meaning of which we are not fully aware (Bargh et al. 2012). If true, these
findings may show that our actions are not as fully consciously determined as
we thought. Some have called into question, however, whether these findings
can be replicated (Chabris et al. 2019), and one could argue that the relevant
cues are merely unattended but not truly subliminal.
Another line of studies focuses on the well-known Libet clock paradigm

(Libet et al. 1983), in which subjective estimates of action onset and conscious
intentions (i.e., the time span during which individuals feel that they are about
to make an action) are reported. The actions concerned are typically simple
motor movements, such as spontaneous flexing of the wrist or pressing of a
button. Preceding these simple movements, it has been reported that the brain
activity arises well before the onset of conscious intention (Deecke et al. 1969;
Libet et al. 1983; Lau et al. 2006; Soon et al. 2008). These findings have stimu-
lated many debates. Simple computational models have also been proposed,
suggesting that the findings are exactly what we should expect since neuronal
processing is noisy (Nikolov et al. 2010; Schurger et al. 2012).
The conclusion from these studies seems to be that our conscious intentions
are preceded by unconscious brain activity. Some have taken this to mean that
our conscious intentions are “determined” by the preceding unconscious activ-
ity, yet this interpretation is unwarranted. In most cases, what was shown is
simply a weak statistical correspondence between the unconscious activity and
the subsequent intention. In any case, whether conscious intentions are deter-
mined is beyond the scope of our current interests.
The Libet studies also led to another finding: the time between our con-
scious intention and actual action execution may be too small for meaningful
causation to take place (Lau et al. 2006). Further, based on early neuroimaging
studies (Lau et al. 2004), the corresponding putative “intention” areas of the
brain can be targeted with magnetic stimulation (Lau et al. 2007): this showed
that the reported onset of conscious intention can be influenced by stimulation
even after the action is completed. Subsequently, other studies have also used
psychophysical methods to produce similar results (Banks and Isham 2009).
Just because intentions may be subsequently revised, however, does not
rule out the possibility that pre-revised versions of intention may occur prior
to action. Ultimately, the actions involved in the Libet studies are simple and
inconsequential, so that even if conscious intentions do not cause them imme-
diately, this does not rule out that intentions may be important for subsequent
and more complex behavior.
Importantly, Wegner and colleagues conducted studies outside of the con-
text of the simple Libet paradigm, concluding that our conscious intentions
may be generally illusory; that is, not causal to immediate actions (Wegner and
Wheatley 1999; Wegner 2002). Likewise, it has been shown that unconscious
information can influence more complex tasks, such as preparing to answer
one type of question over another (Lau and Passingham 2007; Rahnev et al.
2012) or response inhibition (van Gaal and Lamme 2012).
In other studies, it was shown that unconscious information in the brain can
facilitate different forms of associative learning (Taschereau-Dumouchel et al.
2018b), which in some cases revealed therapeutic potential. For instance, using
a technique called multivoxel neuro-reinforcement (Watanabe et al. 2017), one
can pair unconsciously the representations of a spider with monetary reward so
that the subject will subsequently show reduced physiological threat-related re-
sponses to images of spiders (Taschereau-Dumouchel et al. 2018a). In other stud-
ies, using a similar experimental setup, powerful forms of reward-based learning
have been shown to take place unconsciously (Cortese et al. 2019). In addition,
subjective confidence regarding one’s ability to discriminate certain stimuli can
also be changed with this unconscious association method (Cortese et al. 2016).
In summary, unconscious cognitive processing seems very powerful indeed,
especially regarding its ability to form statistical associations and to influence
subsequent behavior (via priming). Does this mean that consciousness plays
no causal role and has no function? The answer is not so straightforward. Lau
(2009) argues that although unconscious processes are powerful, this does not
mean that there is necessarily no room for consciousness to add further ben-
efits. To discuss meaningfully the theoretical possibilities of the role for con-
sciousness, we need to distinguish between different notions of consciousness.
Pure “Qualia” versus Sheer Cognitive Power
In the studies mentioned above, unconsciousness typically refers to stimuli

or processes of which the subjects are unaware; that is, subjects do not know
that they take place. So, in a sense, consciousness is just the opposite: subjects
are aware of the relevant events. There is, however, a tradition in philosophy
that analyzes consciousness as concerning pure qualitative experiences (Nagel
1974; Levine 1983; Block 1995). In some cases, philosophers invite us to con-
sider that molecule-by-molecule functional duplicates of ourselves may lack
such qualitative experiences altogether (Chalmers 1996). While such concep-
tual possibilities are intriguing, this notion of consciousness is not particularly
interesting for our current purposes (Levy 2014). If we define consciousness
as having no functional consequence, of course, it could play zero causal role.
On the other end of the spectrum, one could characterize consciousness
as the capacity to access information and use it for purposeful behaviors. On
one view, this form of consciousness, called access consciousness, is always
supported by strong, stabilized signals broadcast throughout the different sys-
tems in the brain (Dehaene 2014; Dehaene et al. 2017). With this notion of
consciousness, it is highly likely that consciousness will be causally relevant
for important decisions in everyday life (Levy 2014). Once again this seems
to depend, somewhat circularly, on the definition of consciousness we choose
to adopt. Of course, if consciousness is by definition characterized by global,
complex, and elaborate processes, it is likely functionally important. But does
this ignore how much we seem to be able to accomplish unconsciously?
A Middle Ground?
Because of the above considerations, a notion of consciousness that is

relevant for our current analysis should ideally allow for some functional
consequences in principle, without assuming so from the outset (Rosenthal

2012). Is it possible that there is a view that mental event is conscious in
the phenomenological sense while whether it contributes functionally to our
rational thinking and agency ascriptions remains an empirical matter? One
such possible account is the higher-order view of consciousness, which can
be traced back at least to John Locke and Immanuel Kant (Lau and Rosenthal
2011). According to this higher-order view, mental representations of events
in the world are by themselves unconscious. We can call these first-order rep-
resentations. They only become conscious when they are meta-represented
by higher-order representations. That is, higher-order representations about
the first-order representations are necessary for making the content of the
latter conscious. On this view, we can see why powerful forms of uncon-
scious processing are possible: the same first-order representations with the
same functional capacities can be conscious or unconscious, depending on
the presence of the relevant higher-order representations (upon which one
forms beliefs and more complicated narratives). Yet why do we need higher-
order representations for first-order representations to become conscious?
Traditionally, the arguments come from philosophical analysis (Rosenthal
2004): if one is aware of being in a certain mental state, one must represent
oneself as being in that state (via the higher-order representations). Below
we will elaborate further what this means in terms of cognitive architecture.
Criticisms of this well-known theory, and their replies, have been extensively
reviewed (Rosenthal 2005; Brown et al. 2019b).
Concordance with Current Science
Just because a philosophical notion of consciousness exists and can serve our
purpose does not mean that we should adopt it. Fortunately, there is consider-
able empirical support for the higher-order view of consciousness (Lau and
Rosenthal 2011). Neuroimaging studies have shown that subjective aware-
ness of visual stimuli is associated with brain activity in high-level cognitive
regions (e.g., in the prefrontal cortex), even under highly controlled experi-
mental conditions in which the subjects are not merely processing the stimuli
in simple (first-order) tasks (Lau and Passingham 2006). Also, neurological
patients with selective damage of their visual cortex may lose the relevant
subjective visual experiences but not their ability to correctly “guess” the
identity of the relevant stimuli (Weiskrantz 1997); when visual stimuli are
presented to intact parts of their visual cortex (as compared to the “blind”
regions) leading to a conscious experience, higher activity in the prefrontal
cortex was also found (Persaud et al. 2011). Disruption of activity in this
prefrontal brain region selectively impairs one’s ability to introspect whether
one has successfully perceived the stimuli, without impairing (first-order)
perception itself or memory (Rounis et al. 2010; Fleming et al. 2014). These
findings are somewhat difficult to explain if we assume that consciousness is

just strong, global information processing.
While these findings are reviewed in detail elsewhere (Lau and Rosenthal
2011), of particular interest to neuroscience is our emerging ability to assess,
on a person-by-person basis, the efficiency of the relevant higher-order mecha-
nisms (Baird et al. 2013b; McCurdy et al. 2013; Vaccaro and Fleming 2018).
As we will see below, analysis of such individual differences is likely the key
to understanding breakdowns of agency and responsibility.
Another line of support of the higher-order view comes from modern stud-
ies of artificial intelligence (Lau 2019). It is generally accepted that for neural
network models to perform well, they can benefit from having the capacity for
“predictive coding,” in which a model can generate exemplar images (e.g., of
cats or monkeys) top down. This improves the model’s ability to classify im-
ages (e.g., of cats or monkeys). Training such a generative network, however,
can be time consuming. To accelerate this process, another network called the
discriminator, whose job is to detect forgeries (Creswell et al. 2018), can be
trained to discern whether an image is genuinely from the world or created by
the generative network (forgery). When we pit these two networks against each
other, so that the discriminator wins a point for correctly identifying a forgery
and the generative network wins a point by getting away with it, these two
networks grow together not unlike rivaling siblings: they both become highly
efficient in a relatively short time.
In the context of the human brain, it has been suggested that similar mecha-
nisms of predictive coding occur. Neurons in the visual cortex may fire when
a cat is presented to the subject, but similar neural representations are also
involved when we imagine or remember a cat. The brain must be able to tell
what causes these same neural representations to be activated in each instance.
Given that these top-down generative mechanisms seem to be so efficient, it
is likely that through development they are aided by the existence of a dis-
criminator-like mechanism. Such a mechanism can determine whether a visual
representation is generated by oneself or triggered externally by actual objects
of perception. Plausibly, this mechanism can also tell when the same visual
neurons may just be firing because of spontaneous noise. This conceptualiza-
tion fits well with the higher-order view, in the sense that the output of this
putative discriminator is akin to the higher-order representations necessary for
conscious perception to occur. Normal conscious perception happens when the
discriminator decides that a certain visual representation is truthfully repre-
senting the external world right now (Lau 2019).
Formation of Rational Beliefs

In part, based on the findings concerning the surprising power of unconscious
processing, it has been argued that higher-order representations may have little
additional utility besides making the relevant first-order content conscious

(Rosenthal 2012). Using the computational interpretation above, however, we
can identify some plausible key functions. In particular, one such function
may be the formation of rational beliefs. By rational, we merely mean that
these beliefs are subjectively justified in the sense that upon introspection, it
should seem to the subject that the beliefs are reasonable. In general, it seems
reasonable to believe in what we consciously see. Even when we have other
reasons to believe that our perceptual system may be at fault—such as dur-
ing lucid dreaming (Baird et al. 2019), at a live magic show, or after having
knowingly ingested hallucinogens—there is still a strong temptation for us to
believe what we see. According to the above interpretation of the higher-order
view, this is because when we consciously see a cat, we have a first-order
representation of a cat as well as a higher-order representation which states
that the relevant first-order representation is a truthful representation of the
world right now. The derivation from there to the belief that “there is a cat” is
akin to a matter of syllogistic inference, so naturally such a belief may seem
rational from the outset.
What good is having these subjectively justified beliefs? We do not deny
that some beliefs may be unconscious,1 but as human agents, we form narra-
tives about the world and ourselves, and such narratives matter for our actions
(Dweck 1999). In doing so, we tend to try to be coherent (Holyoak and Powell
2016). When we have many beliefs (including background, unconscious be-
liefs), this coherence is often difficult to achieve. One would therefore do well
to include in this rational thinking system only beliefs of which we are reason-
ably certain. According to this perspective, beliefs that enter our narratives are
mostly subjectively justified. In the case of perception, such beliefs can be as
simple as “there is a cat.” However, we also form beliefs about ourselves, our
actions, and emotions, to which we now turn.
Self-Narratives in Agency and Emotions
As in perception, we know that motor representations in the brain also serve

multiple purposes. Simple motor commands in the brain (in the primary and
secondary motor cortices) are activated when we act as well as when we imag-
ine performing the same action or when we observe others performing the
same action. So presumably, as in the case of perception, some discriminator-
like monitoring mechanism needs to decide when a motor command reflects
what oneself is intending to perform (rather than just what one is imagining,
or noise). This allows one to form the corresponding belief that “I intend this
to happen.”
1
What we are claiming here is that beliefs based on conscious experiences are subjectively
justified. Beliefs can, however, be based on unconscious experiences, but they will not be
subjectively justified.
Taking an analogy from computers, the ability to form such beliefs seems
useful. Let us say that a computer server in a large network sends a command
to print ten pages to the printer. Another node in the network complains that the
printer queue is now jammed and asks the first server to resolve the problem. It
would be quite useful for the first server to know who contributed to the print-
ing in the first place.
This self-directed nature of the corresponding representations is also rel-
evant in emotions. Again, we know that for certain basic emotions, imagining
them activates similar neural representations as experiencing them (Reddan
et al. 2018). Likewise, when we emphasize and think about others’ emotions,
similar representations are involved. Therefore, when these first-order emo-
tional states occur, the brain needs to know what the causes are. According
to the view advocated here, one consciously experiences emotions when the
relevant higher-order state points out that the first-order emotional representa-
tion reflects what one is going through. Thus, the corresponding belief may
be simply that “I am angry” or “I am scared.”2 Having such beliefs may be
quite useful in navigating social situations and in explaining to others why we
behave a certain way.
Contrasting these with the relatively simple beliefs in the case of perception
(“there is a cat”; see Figure 12.1), there is a sense that agency and emotions
are intrinsically more self-involving. In fact, as Ledoux argues, without some
minimal concept of the self, an animal may not experience basic emotions
(e.g., fear) at all (LeDoux and Sorrentino 2019).
Why does one need to form these rational beliefs about oneself, which
requires that the relevant first-order states be made conscious via meta-rep-
resentations? The proposal is that first-order processes are powerful: we can
use them to learn about statistical associations between events. However,
mere associations are not coherent narratives. Moreover, narratives are sto-
ries in which events are causally related. When we say that Julius Caesar
invaded a certain country because he was angry, we mean that his emotion
caused certain behavior. When he decided to invade, presumably he saw
himself as an agent who was causally responsible for the decision. Inferring
about causality is, however, notoriously hard. It is technically a challeng-
ing problem from a statistical point of view (Pearl and Mackenzie 2018).
Without controlled experiments in which we can manipulate the putative
causes while holding all other things constant, assumptions need to be made
and heuristics involving counterfactuals may need to be invoked (Bond et al.
2012; Chambon et al. 2018). As such, interpretations matter; there may be
more than one way to tell a story based on the same facts. With these imper-
fect narratives, we form a quasi-rational understanding of why we behave a
certain way, and we provide socially acceptable justifications of our actions,
based on folk psychology.
2
Forming such beliefs does not necessarily involve language ability.
The cat reminds me of lions. There is a cat and there is

Although I know I should not no lion. I feel afraid and
Narrative level OR
be afraid of cats, I still feel want to run. Cats must be
very afraid. So I want to run. scarier than lions??
There is no
I want to
Rational beliefs There is a cat lion; it is just I’m afraid
run
my imagation
Higher-order This reflects This reflects This applies I intend

representation the world my imagination to myself this
First-order Cat Lion Fear Run

representation
Figure 12.1 From first-order representation to self-narrative. Higher-order represen-

tations relate mental states to oneself (e.g., this reflects my imagination) upon which
rational beliefs are formed. Different narratives, however, can be formed depending on
how one relates the beliefs.
Some Related Views
The process of beliefs and subsequent narrative formation that we propose

here is quite similar to the accounts of the post hoc explanations made by
Gazzaniga’s left-brain interpreter (Wolman 2019). To clarify, the above view
does not mean that consciousness is the same as self-narrative (or self-con-
sciousness). The view holds that the necessary and sufficient conditions for
conscious experiences involve having the corresponding first- and higher-or-
der representations. It is only by having these representations that we can form
rational beliefs, based on which we form these causal narratives, while trying
to be as internally coherent as we can.
This view links consciousness with rationality. Therefore, it is related to
other models of rational decision making. In the Two Systems framework,
championed by Daniel Kahneman (2011), first-order representations may
roughly correspond to processes in System 1 (fast), with subsequent processes
belonging to System 2 (slow). In reinforcement learning, there is a well-known
distinction between model-based and model-free learning and decision mak-
ing (Dayan and Berridge 2014). Roughly, higher-order mechanisms may relate
better to model-based processes, whereas the first-order mechanisms may map
to model-free statistical associations.
Our goal is not to replace or compete with these views. They are independently,
empirically well supported, but they may serve different purposes. For example,
although there may be a sense that the Two System approach or model-based
versus model-free distinction may map onto conscious versus unconscious pro-
cesses, such mapping is not intended to be clear cut. The higher-order view, on
the other hand, is a theory of consciousness per se with direct empirical evidence.
In the present context of understanding agency, the consciousness requirement
is an important component of the overall argument. Accordingly, it is not clear if
one should be absolved of the relevant responsibility just because decisions are
made with System 1 (fast) thinking or model-free learning.
Another distinction is that one may conceive of the Two Systems framework
as representing two parallel processes. The first- versus higher-order model
here, however, stipulates that the two mechanisms are in a hierarchy. This hi-
erarchical nature may have important consequences for treatment of mental
disorders; intervention at the first level will causally impact on the higher level
(Taschereau-Dumouchel et al. 2018b).
Understanding Mental Illnesses
This self-narrative account of agency may help us understand why some be-
lieve that patients suffering from severe mental illnesses may be less deserving
of punishment than the unpunctual student, even though in both cases brain ac-
tivity causes the relevant behavior. In the case of a patient with schizophrenia,
the very basic mechanism of higher-order perceptual reality monitoring may
be at fault. Thus, the patient may be unable to distinguish self-generated inner
speech from externally triggered voices (e.g., from “God”), occurring from a
breakdown between the bottom two levels in Figure 12.1. The patient may not
be able to tell if an action is voluntarily produced or controlled by aliens. As
such, the entire self-narrative system may well disintegrate. It is probably not
fair to hold such patients accountable for their own behavior, if they are not
correctly aware of who and what events caused these actions in the first place.
In the case of the unpunctual student, the higher-order system is presum-
ably intact. What might have caused the (mildly) delinquent behavior may be a
momentary overemphasis on the value of not having to rush or an attraction to
another activity. These values represented in the first-order system are no less
brain based and, in a sense, they too cause the resultant behavior. With an in-
tact higher-order system, however, one should be able to appreciate that these
first-order values are problematic and that one would be guilty all the same for
acting a certain way.
What about cases of addiction and substance abuse? Between severe men-
tal illness and everyday cases of delinquency, there likely lies a spectrum. In
some cases of addiction, one may suffer first-order malfunctioning such that a
substance may be associated with an unrealistic expected level of immediate
reward, even when one is well aware that it cannot be good in the long run. In
some cases, this higher-order mechanism may well be relatively intact, so one
may be accountable for not recognizing the situation as such. However, there
may also be cases where such a higher-order system is compromised, due to

chronic abuse, which is known to impair the brain circuitry responsible for
high-level control (Baler and Volkow 2006). Still, even when the higher-order
system is intact, there are cases where alternative actions are not perceived as
feasible or attractive, as suggested by the Rat Park studies (Alexander et al.
1978; Solinas et al. 2008); in such cases, the best solution may well lie in social
policy rather than neurobiology (Hart 2013). Like others (e.g., Levy 2013), to
make the correct judgment we think that we need to be able to adjudicate be-
tween the different cases in terms of both the specifics of the brain impairment
as well as the environment. What we want to emphasize is that the distinctions
between higher-order representations, beliefs, and self-narratives are crucial to
the notion of agency.
A relevant intermediate condition to consider may be obsessive-compulsive
disorders. Here, intrusive thoughts may primarily arise due to a malfunction at
the lower levels; for instance, one may register the scene of an unclean bath-
room as extremely threatening. Among these patients, some may genuinely
believe that this is the case at the self-narrative level. Other patients, however,
may recognize that the unclean bathroom is actually not that harmful, yet the
visceral experience may be too much for them to overcome. In other words,
patients may differ in terms of whether the conscious experience ultimately af-
fects the healthy functioning of the entire higher-order system.
Because of these considerations, we call for more effort to subclassify the
various disorders, including anxiety and depression. Is a certain patient suffer-
ing from malfunctioning at the first-order or higher-level, or both? Importantly,
as LeDoux and Pine (2016) have argued, these different etiologies may need to
be targeted independently. This is complicated by the fact that a disorder at one
level may influence another level, as they are interconnected. To provide com-
prehensive treatment, one useful strategy may well be to target both the higher
and lower levels (DeRubeis et al. 2008). Recognizing which level is the source
of the problem for a particular individual will likely help finesse this process.
Thus, we argue that mental disorders are brain disorders and to under-
stand agency and its breakdown we need to look carefully at the individual’s
condition, using the first- versus higher-order framework. Finding some cor-
relates in the brain for certain misbehavior, however, should not absolve an
individual of relevant responsibility. By understanding which brain correlates
may be reflecting specific behavioral impairments, the framework advocated
here provides a way to identify the theoretically relevant correlates at the dif-
ferent levels.
Understanding Responsibility of Individuals in Society
Mental illnesses can sometimes lead to one of the most devastating conse-
quences: suicide. At times considered to be one of the most personal and
existential decisions (Nietzsche 1955), suicide has also been analyzed as a

social phenomenon (Durkheim 1951). In Emile Durkheim’s classic work on
the topic, he argued that many aspects of suicide depend on societal norms
and values. If suicide were entirely a matter of personal decision or individual
psychopathologies, it would be difficult to account for the stable cross-cultural
differences in suicide rates (Durkheim 1951; Liu 2009). In what ways are we
to understand suicide as being culturally and socially contextualized? Does the
individual not make the decision after all?
The advocated view in this chapter can help us categorize the different ways
in which social influences take place. At the first-order level, the statistical
regularities of social events are picked up by the individual: How rarely does
suicide occur? When the tragic event of suicide occurs, how do others react?
As the individual becomes consciously aware of these events and their contin-
gencies, the individual forms rational beliefs about suicide in social settings.
Importantly, one also forms narratives about these events, in which one as an
agent plays certain causal roles.
At the narrative level, interpretation matters. Different stories have differ-
ent meanings. We interpret victims of suicide as causal agents. Why did so-
and-so kill themselves? What drove them to such a desperate decision? Was it
moral for them to do so? How does it causally affect their loved ones? These
narratives apply to others as well as oneself and are naturally colored by our
social understanding of the relevant concepts and implications. As such, so-
cietal norms and values influence suicidal behavior. However, we ultimately
understand the decision is to be made by the individual concerned, within the
given social context (Weber 1930).
Take racism as another example. One may perceive the presence of many
youngsters of a certain ethnic group in a neighborhood as being statistically
correlated with a higher occurrence of crime. Hypothetically, let us say that
this association were statistically true in a certain context. Our unconscious
first-order system may be truthfully picking up such an association, yet at
the narrative level, where we ascribe causal relations to events, we do not
necessarily have to interpret the ethnicity of the relevant youngsters as the
cause of the prevalence of crime. More plausibly, both the ethnic makeup
and occurrence of crimes in the neighborhood could be commonly caused
by poverty and other forms of social injustice. Accordingly, the individual
is still responsible for making the correct and appropriate interpretation and
forming the correct narrative, given the same statistical facts picked up by the
first-order system.
In this context, it is worth noting that in overcoming racism, much effort
is focused on addressing our various implicit biases. What we have argued
for here (excluding any possible statistical biases) is that there is also a level
of narration to consider, where cause and effect is up for debate. Just because
narratives are subject to interpretation does not mean they are too elusive to
be worth studying. They can be changed and clarified through education and
social discourse. At times, focusing on this higher level may well be more ef-
fective than methods focusing on changing our first-order implicit biases.
Closing Remarks on the Role of Higher-Order

Mechanisms in Intrusive Thinking and Treatment
When considering therapeutic intervention in higher- or first-order frame-

works, it may be useful to consider B. F. Skinner’s perspective on whether
the concept of human consciousness plays any meaningful role in science
(Blanshard and Skinner 1967:325): “No major behaviorist has ever argued that
science must limit itself to public events.…As a behaviorist, however, I ques-
tion the nature of such events and their role in the prediction and control of
behavior” (italicized emphasis added). Skinner famously advocated studies in
psychology based on the Pavlovian tradition, in which we focus on how an
individual learns about the statistical associations between observable events,
rewards, and behavioral responses. In our terms, this concerns the unconscious
first-order level. Skinner (1971) went so far as to speculate that to engineer a
better society, we should focus precisely on these basic mechanisms. Indeed,
in health and disease, methods of intervention based on these Pavlovian prin-
ciples have often been proven effective. However, as we see from the quote
above, even a stern behaviorist as Skinner did not completely rule out the pos-
sibility that consciousness could ever be studied. The problem is that at this
higher narrative level, things are complex and often depend on social factors
that affect the various interpretations. Unlike Skinner, however, we are not so
pessimistic about the possibility that we can understand this system. We have
laid out how higher-order level narratives may be causal and may also inform
ourselves about our role as causal agents. In the context of intrusive thinking,
this means that whether a thought is considered intrusive or voluntary may ul-
timately depend on these imperfect narratives. The narratives themselves may
be complex, depending on perspectives, but at least we can try to understand
the underlying general mechanisms.
To conclude, we have provided some limited evidence, but no proof, that
this framework is correct. The issues at hand are of immense historical sig-
nificance. They concern whether we can understand individuals as rational
agents. Prior to World War I and II, many great scholars from disciplines as
diverse as sociology, economics, political science, psychology, and neurosci-
ence opined on this topic. However, over the last half century, discussions on
consciousness and free will have shifted toward physics. We suspect this may
be due to contingent sociohistorical factors, some unfortunate and not neces-
sarily productive. The important question of freedom of the individual, in the
context of health and disease, in isolation as well as within social networks,
may not benefit as much from insights from physics as from careful analysis of
the neurocognitive structure (outlined throughout this volume) that underlies
our narratives about ourselves. That psychotherapy can be just as effective as

Pavlovian-based methods in the clinic, at least in some cases, suggests that we
should not write off the intriguing possibility that our higher-order mechanisms
can also be systematically understood, and modified, as needed.
13
Systems Approach to
Intrusive Experiences
Angela C. Roberts, Rita Z. Goldstein, David Badre,
Bernard W. Balleine, Hugo D. Critchley,
Aikaterini Fotopoulou, Sophia Frangou,
Karl J. Friston, Tiago V. Maia, and Elliot A. Stein
Introduction
This chapter explores how intrusive experiences may occur at a systems level from psy-
chological, computational, neurobiological, and physiological perspectives. A general
scheme is proposed of the essential elements of an intrusive experience, and where in
this scheme dysregulation could occur to increase the likelihood of an intrusive experi-
ence. It also considers a range of psychological and mathematical models that have
been applied to explain how intrusions may ultimately happen, some of which are more
closely integrated into neurobiological systems than others. These include a Bayes-
ian model of active inference, integrated psychological and physiological models of
interoception, and psychological and neurobiological models of working memory and
associative learning and their relevance to concepts of flexibility and stability.
Phenomenology of Intrusive Experiences
Human mental operations (e.g., perception, emotion, cognition, metacogni-

tion, and action planning) are both complex and diverse. It is therefore im-
portant that we clearly define the phenomenological properties of intrusive
thinking, particularly because it can encompass a wide array of forms, topics,
and themes (see Visser et al., this volume). Here, we employ the term intru-
sive experience instead of intrusive thinking to denote that our deliberations
apply to intrusive verbal thoughts, intrusive nonverbal thoughts (e.g., images,
Group photos (top left to bottom right) Angela Roberts, Rita Goldstein, Bernard
Balleine, Sophia Frangou, Hugo Critchley, Elliot Stein, Tiago Maia, Katerina
Fotopoulou, Karl Friston, Elliot Stein, Bernard Balleine, Rita Goldstein, Hugo Critchley,
Karl Friston, David Badre, Katerina Fotopoulou, Angela Roberts, Tiago Maia, Sophia
Frangou, David Badre, Angela Roberts
246 A. C. Roberts et al.
music), intrusive impulses (e.g., motor actions), as well as intrusive bodily

sensations.
Intrusive experiences have been conceptualized in varying ways
(Rachman and Hodgson 1980; Parkinson and Rachman 1981; Salkovskis
and Harrison 1984; Edwards and Dickerson 1987b; Freeston et al. 1991;
Yao et al. 1999), and an overall consensus definition is currently lacking.
The most common features across definitions involve the involuntary and
disruptive nature and internal attribution of intrusive experiences; with re-
gard to valence and controllability, there is greater variation (Rachman and
Hodgson 1980; Parkinson and Rachman 1981; Salkovskis and Harrison
1984; Edwards and Dickerson 1987b; Moulding et al. 2014). Rather than at-
tempting to provide a general definition of intrusive experiences—a goal that
has tended to elude the field and has been tackled in more detail by Visser
et al. (this volume)—we focus on three stages inherent to intrusive experi-
ences (Figure 13.1). This deconstruction allows for the empirical probing of
the processes and neural systems that underlie intrusive experiences, with
the ultimate goal of identifying the most appropriate targets for intervention
when such experiences become pathological. Consequent upon this model
are the following parameters:
• The intrusion itself is inherently neutral. It is conceptualized here as a
neural event (or cascade of events), the origins of which are likely to be
relatively localized within specific brain circuits or networks.
• Intrusions undergo appraisal. During appraisal, attributes are assigned
to the intrusion. By definition, intrusions are unintended and thus they
will be appraised as involuntary. Assignment of other attributes and
emotional responses to the intrusion will depend on its nature, content,
and context (situational and personal) in which it occurs.
• Post-appraisal cognitive control mechanisms (Braver 2012) determine
the response strategy to the intrusion.
• The resulting intrusive experience is not inherently pathological but
rather a common universal human experience (Salkovskis and Harrison
1984; Freeston et al. 1991; Corcoran and Woody 2008; Bouvard et al.
2017). Some intrusive experiences, however, can be pathological, de-
pending on their nature, content attributes, recurrence, controllability,
and behavioral consequences (e.g., Julien et al. 2007; May et al. 2015).
Intrusion Appraisal Outcome
Figure 13.1 Components of the intrusion experience.
Systems Approach to Intrusive Experiences 247
The Intrusion
An intrusive experience has a neural “locus of origin,” is sufficiently strong so

that it spreads to brain regions with which it is closely linked, and propagates
beyond a critical threshold which allows it to interrupt other processes and
enter awareness.
The locus of origin can provide an intuitive account for the nature and con-
tent of the intrusive experience (Figure 13.2). Intrusive experiences of a sen-
sory nature (e.g., images, music) are likely to originate within sensory systems.
Intrusive experiences that involve movement are likely to originate in motor
systems. Intrusive experiences that involve somatic sensations (e.g., thirst) are
likely to originate in homeostatic systems (Figure 13.2; for a discussion of dif-
ferent neurological intrusion domains, see Gourley et al., this volume).
Tourette syndrome is a good example of where a locus of origin for the
intrusive experiences can be identified. In Tourette syndrome, intrusive pre-
monitory sensations and movements (i.e., tics) are associated with abnormal
activation in somatosensory and motor cortical regions (Conceição et al.
2017). A locus of origin formulation is more challenging for intrusive experi-
ences involving verbal thoughts. Recent advances in cognitive neuroscience,
however, suggest that cognition in everyday life is dominated by thoughts that
are not directly linked to sensory processing or task-directed behavior (Kane
et al. 2007). Several terms (e.g., spontaneous cognition, unconstrained cogni-
tion, or mind wandering) are currently used to refer to these stimulus- and
task-independent processes. In parallel, emerging neuroimaging findings have
associated spontaneous cognition with connectivity within the default mode
network, a functional brain network that is more active during stimulus- and
task-independent periods (Andrews-Hanna et al. 2010; Dixon et al. 2014).
However, it is important to note that our model postulates that regardless of the
initial locus, the originating signals spread to additional brain regions follow-
ing connectivity pathways so that intrusive experiences acquire multisystem
associations once they reach a certain threshold (see below). In other words,
they enter the global workspace (Dehaene et al. 1998) or form part of the win-
ning coalition (Maia and Cleeremans 2005).
To account for how intrusive experiences occur, we propose two heuris-
tic mechanisms: a breach and a permissive mechanism (Figure 13.3). These
mechanisms are described separately although they may coexist. They are
conceptually embedded within theories that view experience as the out-
come of selective signal propagation in the face of competition (Dehaene
and Changeux 2004; Beck and Kastner 2009; Graziano and Webb 2015) or
global constraint satisfaction (Maia and Cleeremans 2005). The mechanisms
for signal selection are currently unclear and have been described with vari-
ous terms, including signal biasing or weighting (Sergent and Dehaene 2004;
Beck and Kastner 2009), signal enhancement (Graziano and Webb 2015), bi-
ased competition (Desimone 1998; Deco and Rolls 2005), and gating (as we
248
IC
IC
dACC
dACC
Amy
TP
Integration
monitoring
systems
A Model of the Brain
cognition system
cognitive system
Goal-directed
Spontaneous
Homeostatic
Sensory
systems
systems
systems
Motor
Hypo
Medulla
PAG
LCC
NTS
Amy
Figure 13.2 The brain is functionally organized into cognitive systems supported by spatially defined networks (Power et al. 2011). This simplified
model illustrates brain activity within brain systems (left) that could host a potential “locus of origin” of an intrusion: sensory and motor systems,
homeostatic systems involved in the integration of exteroceptive and interoceptive signals (Salvato et al. 2020), a system involved in spontaneous
cognition supported by the default mode network (Raichle et al. 2001), and a system for goal-directed behavior supported by frontoparietal regions
(Fox et al. 2005). The model posits that intrusions are experienced as such when they enter “awareness,” which is likely to occur in the presence
of additional recruitment of networks involved in monitoring (shown on the right), such as the salience network (Seeley et al. 2007). The blue ar-
eas (middle) represent the network space, with oscillatory activity therein denoted by red lines. Amy (amygdala), dACC (dorsal anterior cingulate
cortex), Hypo (hypothalamus), IC (insula cortex), PAG (periaqueductal gray), LCC (lateral cerebral cortex), NTS (nucleus tractus solitarius), TP
(temporal pole).
Increased strength
Normal threshold
Reduced threshold
Breach intrusion experiences Permissive intrusion experiences

Increased signal strength Lower awareness threshold
249
Figure 13.3 Breach and permissive intrusive experiences.
discuss in more detail below). Here we use the analogy of awareness thresh-
old (borrowed from sensory perception) to visualize the moment a signal
gains sufficient biological momentum to breach the threshold of awareness.
Accordingly, breach intrusive experiences can occur because the strength,
features, or contextual significance of the originating signal enables its selec-
tive enhancement. In contrast, permissive intrusive experiences occur when
the threshold is transiently or persistently lowered and thus permits the propa-
gation of weaker signals. The timing of intrusive experiences (i.e., when they
occur) can be influenced at any point by the external environment as well as
by internal states, which can be referred to as “motivational states” in that
they combine representations of somatic states and overall general behavioral
drives (discussed in more detail below). It also follows that intrusive experi-
ences are influenced by genetic and molecular factors, including neurotrans-
mitters (e.g., Bonvicini et al. 2016; Sinopoli et al. 2017), that define healthy
within-individual variation (i.e., the likelihood of an intrusion within an indi-
vidual) and interindividual differences (i.e., differences between individuals
in the likelihood of experiencing intrusions), and that these may be associated
with pathological conditions affecting brain integrity at multiple organiza-
tional levels (e.g., Keelan et al. 2019).
Generally, signals relating to survival (e.g., hypoglycemia) will generate
breach intrusive experiences. The same could apply to abnormally generated
signals, as in the case of Tourette syndrome, where abnormal sensorimotor ac-
tivation spreads to other brain regions (e.g., the insula) and eventually breaches
the threshold of awareness (Conceição et al. 2017). Signals relating to signifi-
cant prior (e.g., childhood abuse, traumatic event) or immediate circumstances
(e.g., negative thoughts about the self) may also be selectively enhanced and
thus breach the awareness threshold. In such cases, the content of the intrusive
experiences is more likely to be “personal” to the individual. The personal
nature of the intrusion is also likely to constrain the range of its content; thus,
such intrusive experiences are likely to be stereotypical. The intrusion experi-
ences observed in posttraumatic stress disorder (PTSD) are prime examples as
their content is repetitious and of personal significance (American Psychiatric
Association 2013). Our model also predicts that permissive intrusive experi-
ences are likely to have a more variable and circumstantial content because the
lowering of the awareness threshold will permit the propagation of a variety of
signals. Attention deficit hyperactivity disorder (ADHD) would be a prototypi-
cal example of a condition in which permissive intrusive events might occur.
Currently, intrusive experiences in ADHD are considered in terms of abnor-
malities in attentional brain systems that gate awareness (Castellanos and Proal
2012; Bozhilova et al. 2018). As already mentioned, the dichotomization of
intrusive experiences as breach or permissive does not imply that they are mu-
tually exclusive. For example, up to 50% of patients with Tourette syndrome
have ADHD, suggesting that breach and permissive intrusions may co-occur
and determine clinical severity and complexity.
The Appraisal
During the appraisal stage, we postulate that the intrusive experience will at-
tract unconditional and conditional attributes and emotional states. By defini-
tion, intrusive experiences will be unconditionally labeled as involuntary as
they bypass processes of agency (see Liu and Lau, this volume; Gallagher
2012; Moore and Fletcher 2012; Braun et al. 2018). However, typical intrusive
experiences retain the “sense of ownership”; that is, the sense of selfhood we
attribute to our own bodily sensations, thoughts, and actions (Gallagher 2012).
It is worth noting that they are distinct from psychotic experiences which, al-
though often construed as intrusive (especially hallucinations and delusions),
typically involve a loss of agency and self-ownership (Feinberg 1978; Moore
and Fletcher 2012; Frith 2014).
The appraisal of intrusive experiences is a multisystem phenomenon that
may, in some cases, rely on complex representations involving semantic/
linguistic networks. During appraisal, the attributes assigned to intrusive ex-
periences and the emotional responses they invoke will depend on their con-
tent, nature, and normative significance (i.e., alignment of personal beliefs
and societal values) (Korsgaard 2009). We argue that the ultimate purpose of
the appraisal is to determine the “likedness” of the intrusive experience; that
is, the degree to which the experience is aligned with the individual’s future
plans (Figure 13.4). As used here, likedness aligns with notions of motivational
relevance (Higgins 2011) and self-congruence (Rogers 1959; Higgins 1987)
and, as mentioned above, the appraisal of the intrusive experience depends
on the characteristics of the individual having the experience, including their
exposures.
Appraisal
Unconditional labeling Conditional labeling

(depends on content, nature, context)
Involuntary Variable attributes

(e.g., distressing/pleasant;
ego dystonic or syntonic)
Liked Unliked
Help
seeking
Figure 13.4 Appraisal of intrusive experiences.
Intrusive experiences that are appraised as distressing and “not liked” are
more likely to be classified as clinically significant and to elicit help-seeking
behavior. However, intrusive experiences can be of a positive nature and
liked, as in thoughts associated with loved ones or that emerge from sud-
den insight or “eureka” moments (Kounios and Beeman 2014). Still, intru-
sive experiences that are deemed positive are not always adaptive and may
contribute to further pathology by providing confirmation for maladaptive
beliefs, as in hedonic hunger in individuals with restrictive eating disorders
(Lowe et al. 2016).
The Outcome
The outcome of the appraisal will invoke mechanisms and networks that sup-
port selective attention, decision making, response inhibition, and response
selection (Niendam et al. 2012; Langner and Eickhoff 2013; Zhang et al. 2017;
Chen et al. 2018b). We assume that there will be no voluntary inhibition for
liked intrusive experiences (Figure 13.5). The experience would either be al-
lowed to decay or it could be maintained through attentional mechanisms. A
liked intrusive experience may even act as a catalyst or starting point for an-
other mental or motor plan. In such cases, the switch from the pre-intrusion
state to a new one may be viewed as a positive outcome of the intrusive events.
Eureka moments would fall under this category.
By contrast, “unliked” intrusive experiences will evoke attempts at volun-
tary inhibition. The success or failure of the experience will depend on the
functional integrity of frontostriatal networks that are generally implicated
Outcome
Liked
Unliked
Voluntary
Inhibit
perseveration
Liked Unliked
Switch to new Involuntary
process perseveration
Help
Normal
seeking
Figure 13.5 Outcome of intrusive experiences.
in inhibitory control (Niendam et al. 2012; see also chapters by Balleine and
Badre, this volume; Bari and Robbins 2013). Of note, disorders characterized
by intrusive experiences also present with a more general impairment of inhib-
itory control that affects multiple aspects of cognition and behavior (Gourley
et al., this volume; Marsh et al. 2009; Shin et al. 2014; Morand-Beaulieu et al.
2017; Pievsky and McGrath 2018). Failure of inhibitory control is expected
to give rise to perseveration and/or premature action (as exemplified in com-
pulsivity and impulsivity, respectively), which may elicit secondary appraisals
involving frustration, anger, and increased arousal directed at the failure to
inhibit rather than the original intrusive experience. Such an outcome is likely
to increase the allocation of attentional resources to the intrusive experience
and the inhibitory failure; in some individuals, this may reinforce intrusion
experiences, leading to a pathological loop.
Salience, Precision, and Value in Intrusive Experiences
Having considered the nature of intrusive experience in terms of definitions,

phenomenology, and their implications in a clinical setting, this section pro-
vides a complementary perspective that takes its lead from systems neuro-
science and, in particular, computational approaches that offer a formal and
quantitative account of the phenomenology at hand. We introduce concepts
of precision, salience, and (motivational) value that may help understand how
and why intrusive experiences occur. We use two working examples that il-
lustrate how dysregulation within these psychological domains may explain
different sorts of intrusive experiences; namely, those associated with obses-
sive-compulsive disorder (OCD) and PTSD. This section concludes with a dis-
cussion of the conceptual implications in terms of computational architectures
that underwrite intrusive experience, and how accompanying computational
models and (neuronal) process theories can be used to characterize empirically
observed behavioral and neuronal responses.
Brief Review of Active Inference with a Special Focus on the

Nature of Precision, Salience, and Value
The treatment in this section considers cognition as a process of inference or

belief updating in the brain. Specifically, we use an active inference frame-
work to cast action and perception as solving an inference problem; namely,
optimizing beliefs about states of affairs in the lived world and, crucially, be-
liefs about how the world should be sampled or navigated (i.e., beliefs about
plans or actions). In short, we make a simplifying assumption that trains of
thought can be associated with planning as inference (Attias 2003; Baker et
al. 2009; Botvinick and Toussaint 2012; Baker and Tenenbaum 2014; Mirza
et al. 2016).
Planning as inference rests on an internal model of how (unobserved)

states of affairs causing (observed) sensations are generated. This is known
as a generative model, usually expressed mathematically in terms of the
likelihood of some observations, given latent or hidden states and prior
beliefs about those states (for a more detailed account, see Appendix 13.1
and Figure 13.A1). In this setting, beliefs are nonpropositional (i.e., sub-
personal) and simply refer to probability distributions encoded by synaptic
activity or connectivity (in the sense of Bayesian belief updating or belief
propagation). A simple example of active inference is the way that we for-
age for visual information. If I move my eyes from one position to another,
the state of my oculomotor system will change, and this will have profound
implications for the sensory impressions on my retina. A sequence of eye
movements would then correspond to a particular policy or action strategy.
My job is to infer the most likely policy that “something or someone like
me” would engage, and then select a particular action (i.e., a next move)
under that policy.
In selecting the most likely policy, I will necessarily refer to my prior be-
liefs about the policies I am likely to pursue; namely, those that provide the
most evidence for my (generative) model of the world. This can be expressed
formally in terms of a prior over policies, based on expected free energy. Free
energy, in this instance, is known as an evidence bound in machine learning
(Winn and Bishop 2005) and can be thought of as a measure of expected sur-
prise or prediction error. Mathematically, expected surprise is also known as
uncertainty. This means that I will select those policies (and implicit courses of
action) that resolve uncertainty about the state of the world. This formulation
of active inference emphasizes the two-way exchange between an agent and
her world, where the implicit action-perception cycle means effectively that
beliefs can change states of the world, which in turn change the sensations that
update beliefs. For an illustration of this circular causality, see Figure 13.A2
in Appendix 13.1.
Motivational Value and Salience
Mathematically, expected free energy can be decomposed in a number of ways

(see Figure 13.A2 for a decomposition into risk and ambiguity). For our pur-
poses, the more prescient decomposition is in terms of salience and value.
Heuristically, the (negative) expected free energy of a policy is equal to sa-
lience plus value (see Appendix 13.1 for details and how this decomposition
relates to other disciplines in neuroscience):
Expected free energy = Salience + Expected value. (13.1)
In this setting, salience corresponds to the uncertainty resolving or intrinsic
(epistemic) value of a policy. It is variously referred to as relative entropy,
mutual information, information gain, Bayesian surprise, intrinsic motivation,
or value of information (Barlow 1961; Howard 1966; Optican and Richmond

1987; Linsker 1990; Itti and Baldi 2009). Salience, therefore, reflects the infor-
mation gain or resolution of uncertainty afforded by response to a cue: “How
much will I learn, if I look over there?”
Salience can be contrasted with expected (extrinsic or instrumental) value,
which is the motivational value of a policy defined in terms of outcomes that
are preferred a priori. Expected value is an important construct in optimal con-
trol theory in engineering, reinforcement learning in psychology, and utility
theory in economics. Expected value simply scores the expected returns (cf. re-
wards) following a particular policy, expressed in terms of the log probability
of some prior preferences (i.e., preferred or expected outcomes). This is some-
times referred to as extrinsic value, as opposed to epistemic value, to make it
clear that these are extrinsically supplied outcomes that provide a motivational
value for the policies under consideration. The foregoing offers a definition of
salience and motivational value in terms of active inference and the accompa-
nying quantities or functionals of Bayesian beliefs encoded by neuronal activ-
ity and connectivity. So, what about precision?
Precision and Attention
Precision is an attribute of sensory outcomes or evidence at hand. Very precise

data are informative, in the sense of having a high signal to noise. The impor-
tant thing, from our perspective, is that precision has to be estimated or inferred
in a context-sensitive fashion. For example, if I know that I am exploring an
unfamiliar room in the dark, I know that the precision of visual sensations will
be much lower than the precision of my somatosensory sensations. I would
therefore assign a greater precision to the mapping between the hidden states
of the world (e.g., “a chair in front of me”) and the somatosensory outcomes
(e.g., “I will feel this chair if palpated”). Conversely, if I know the light is on,
I will adjust the precision of my visual mapping such that visual information
is afforded much more precision and has a much greater influence on belief
updating about the state of my room.
Psychologically, this is effectively the same as attention (Desimone et
al. 1990; Desimone 1998; Womelsdorf et al. 2007; Parr and Friston 2019);
in other words, a selective gating or attentional filtering affords one sort of
sensory stream with more precision than another. When this deployment
of attention is part of a policy (i.e., a covert action much like the premo-
tor theory of attention), we have an attentional policy that is implemented
through a selective gating of the sensory information at hand (Limanowski
and Friston 2018). This will become an important concept later in our dis-
cussion of memory (see below), where policies that selectively afford pre-
cision to different sources of information correspond to gating policies. In
hierarchical generative models, the level of the implicit gating or precision
control may determine the nature of attention: exogenous versus endogenous
(e.g., spatially directed attention vs. attention to a particular visual feature,

respectively).
Applying Computational Models of Active Inference to Understand

Intrusions in Obsessive-Compulsive Disorder and Posttraumatic
Stress Disorder
According to the above formulation, precision is a ubiquitous attribute of the
likelihood and prior beliefs that nuance or select the right kind of information
for belief updating. This selection or gating rests on the excitability or lat-
eral inhibition among competing representations at any level of a hierarchical
generative model. As noted above, precision is context sensitive and must be
inferred; this means that it depends on beliefs about hidden states (i.e., con-
text) and, indeed, beliefs about policies (i.e., “what I am doing”). In contrast,
salience and value are attributes of a particular plan or policy whose evalua-
tion involves belief updates about the succession of states in the future, under
a particular course of action. A salient act is one that resolves uncertainty or is
likely to have the greatest epistemic affordance. The value of a plan is scored
by the degree to which the outcomes are likely to be realized.
Let us now consider the computational pathology that might underwrite
a typical intrusive experience in OCD: “checking behavior.” Assume that
there are two states of the world with which I am concerned: “the door is
locked” versus “the door is unlocked.” Any policies that resolve uncertainty
about whether the door is locked will have a high salience. If I do not know a
priori whether the door is locked or not, checking whether the door is locked
has the greater salience and will, in the healthy course of things, resolve my
uncertainty.
Imagine now that my generative model also predicts a state of physiologi-
cal arousal due to the possibility that the door is unlocked, and all the cata-
strophic consequences that such a state of affairs could entail. If I can resolve
my uncertainty and be 100% certain that the door is locked, then I predict that
the associated interoceptive evidence for physiological arousal will also be
attenuated. Now, imagine what would happen if I were unable to attenuate the
precision of interoceptive signals:1 I would check the door, expecting to find
it locked and expecting my arousal to subside, but it does not.
I would now be in the curious situation of still being uncertain about when
the door is locked because I have sensory (interoceptive) evidence at hand that
I cannot have checked the door (because I am still physiologically aroused).
This means that the epistemic affordance of door checking is still in play.
In fact, unless I can attenuate my interoceptive signals, this uncertainty will
1
In active inference, a failure to attenuate the precision of proprioceptive or interoceptive sig-
nals is accompanied by a failure to engage motor or autonomic reflexes. In this example, a
failure to engage autonomic reflexes means that a state of physiological (sympathetic) arousal
would persist.
continue to be in play and induce successive checking behavior that may pro-
ceed indefinitely.
Notice in this example that checking behavior has been formulated in terms
of aberrant salience because the action of rechecking the door does not lead to
the resolution of uncertainty. This aberrant salience is suboptimal (i.e., patho-
logical) because of a failure to attenuate or change interoceptive signals. In
short, a failure of sensory attenuation led to aberrant salience and a persistent
epistemic affordance that never resolves itself. In other words, no matter how
many times I check the door, I never sense that my uncertainty has been re-
solved, which could further maintain a state of autonomic arousal. Expressed
even more simply, this checking behavior is futile because there is an irre-
ducible uncertainty about the state of the world due to a failure to attenuate
interoceptive evidence from my body. This predictive processing, or active
inference account of OCD, is based (and elaborates) on work by Kiverstein et
al. (2019) and Rae et al. (2019a), and owes much to seminal accounts of why
patients with OCD appear to be “stuck in a loop.”2 For example, Roger Pitman
(1987:336) suggested that “the core problem in OCD is the persistence of high
error signals, or mismatch, that cannot be reduced to zero through behavioral
output,” and that “the obsessive-compulsive’s internal comparator mechanism
is faulty. No matter what perceptual input it receives, it continues to register
mismatch….It may be that in fact the action was well done, but the defective
comparator cannot register it” (Pitman 1987:340).
In turn, Szechtman and Woody (2004:111) suggest that “the symptoms
of obsessive-compulsive disorder...have what might be termed an epistemic
origin—that is, they stem from an inability to generate the normal ‘feeling
of knowing’ that would otherwise signal task completion.” On the empirical
side, Gentsch et al. (2012:656) found decreased sensory attenuation in OCD,
which was suggested to “explain the tendency of individuals with OCD to con-
tinuously register error signals, and to experience dissatisfaction in outcome
processing.”
The somewhat contrived formulation of OCD, in terms of aberrant salience,
focused on an account of intrusive experience that manifests in overt motor
behavior. Does this explanation hold for intrusive thoughts, images, and expe-
riences in PTSD? A plausible account could proceed along the following lines:
Imagine that, at the point a traumatic event is experienced, there is some par-
ticular configuration of (interoceptive or exteroceptive) sensory inputs in play.
The traumatic event can then induce a one-shot learning of the concomitant
gating policy. When this pattern of sensations is encountered subsequently, it
is extremely difficult to ignore, because sensory information is afforded great
precision. These sensory cues will induce belief updating and the selection of
2
This account is from the PhD thesis by Itzchak (Isaac) Fradkin: “Deficits in processing of
prediction errors in obsessive compulsive disorder: Effects on action, thoughts, learning and
agency,” Hebrew University of Jerusalem, June 2019.
the traumatic narrative or policy that entails overt or covert action. In the latter
setting, action is neither motoric (i.e., mediated by striated muscles) nor auto-
nomic (mediated by smooth muscles) but attentional in nature. In other words,
the gating policy is called up in an obligatory fashion, sometimes described in
terms of modulating sensory and prior precision (Skewes et al. 2014; Ainley et
al. 2016; Powers et al. 2017; Rae et al. 2019a).
This traumatic active inference or learning will induce a recapitulation of
the internal policy or narrative that may enable posterior expectations all the
way down to the sensory levels of perceptual hierarchies. In other words, a
triggering event will breach attentional thresholds and induce a cascade of
hierarchical and sequential processing that recapitulates the sequential narra-
tive associated with the original trauma. The mechanisms behind such fictive
(intrusive) experience are part and parcel of self-evidencing under a generative
model. Common examples here include dreaming, imagination, and the gen-
erative or constructive perceptual processing associated with structure learning
and eureka moments (Hinton et al. 1995; Botvinick et al. 2009; Gershman and
Niv 2010; Tervo et al. 2016; Friston et al. 2017; Gershman 2017).
Based on this account, the intrusive experience induces a gating policy
that prescribes covert (mental) actions that are manifest as internal scene con-
struction and accompanying narratives (Peters et al. 2017; Wilkinson et al.
2017), as opposed to the mostly overt actions considered in the OCD example
above. Clearly, the foregoing account does not offer a qualitative distinction
between intrusive experiences that reflect an adaptive response to trauma and
the psychopathology that results when intrusions are experienced (or manifest)
as maladaptive and persistent. However, the computational account narrows
down the field, in terms of where aberrant inference and learning may be oper-
ating in conditions like OCD and PTSD. Next, we consider the failure of sen-
sory attenuation and subsequent failure to relearn the right sort of attentional
response as a plausible candidate.
Summary
The two working examples of OCD and PTSD were introduced here to make
a key point: the intrusive experience of OCD rests upon aberrant salience
that is secondary to a failure of sensory attenuation; namely, an aberrant top-
down modulation of sensory mappings. In contrast, the PTSD example appeals
only to aberrant precision via a breach of sensory attenuation due to traumatic
learning of a particular attentional set or gating policy. In other words, people
with PTSD may lose the capacity to ignore the irrelevant and be plagued by
breaches of attentional filtering or gating endowed by sensory attenuation. If
one subscribes to these accounts, the conclusion is that the primary pathophys-
iology behind both kinds of intrusive experience is a failure of sensory attenu-
ation that most likely involves interoceptive signals. Interestingly, a failure of
sensory attenuation emerges in computational treatments of other psychiatric
conditions (Skewes et al. 2014; Ainley et al. 2016; Powers et al. 2017; Rae et
al. 2019a); in particular, schizophrenia and autism. [For a review of aberrant
precision and sensory attenuation in psychiatry, see Stephan et al. (2016) and
Friston (2017) for details and references.]
On this account, a minimal but sufficient explanation for intrusive experi-
ence is a failure of inhibitory control inherent in the sensory attenuation. The
key thing that the active inference framework brings to the table is that this in-
hibitory control is not about the contents of perceptual experience, but the pre-
cision or attention afforded this content. From a physiological perspective, this
is important because a failure of inhibition (i.e., a failure of sensory attenuation
or attenuation of sensory precision) may be mediated not by hyper- (or de-)
polarizing neuronal populations but by modulating their excitability or gain.
In turn, this suggests the mechanisms that underwrite the pathophysiology of
intrusive experiences are located either in classical modulatory neurotransmit-
ter systems or the downstream effects on cortical excitability (as mediated by
fast-spiking inhibitory interneuron coupling with pyramidal cells).
In summary, the emerging picture is of a deficit in the neuromodulatory
mechanisms (and dynamics) that implement the top-down control of atten-
tion; namely, its sensory attenuation. A natural corollary is that there may be
as many different forms of intrusive pathologies as there are neuromodulation
mechanisms and projections. Irrespective of this diversity, and the accompa-
nying regional specificity of evidence accumulation schemes in the brain, one
underlying mechanism becomes apparent: the breach of sensory attenuation
(i.e., attentional filtering) by exogenously or endogenously generated cues that
underwrite belief updating about states of the world and our active engagement
with that world. Clearly, in many instances, this intrusion is part of normal per-
ceptual synthesis and subsequent planning. For example, a loud noise is salient
because it offers a person the opportunity to “look over there” and resolve any
uncertainty associated with the surprising sensory signal.
The pathology implicit in the examples above rests on aberrant salience
that maintains irreducible uncertainty incurred through a failure to attenuate
interoceptive signals (as in the case of overt compulsive behavior in OCD).
It can also rest on the failure of sensory attenuation to be attributed to, and
subsequent failure to relearn, the right kind of attentional response to triggers
(as in the case of PTSD). As discussed above, the notion of a breach in sensory
attenuation is a key aspect of higher-order models of intrusive experiences that
consider the evaluation (i.e., the appraisal) of inferred states and subsequent
metacognitive influences. At present, three conclusions follow from the formal
analysis of this section that are remarkably consistent with the treatments of-
fered in other chapters in this volume:
• Intrusive experiences are inherently interruptive in the sense that they
induce a quantitative change in the selection of narratives or sequential
policies, which underwrite overt or covert (mental) action.
• These intrusive episodes (events) are experienced in virtue of being

manifest in terms of beliefs about (overt or covert) action. This follows
because there is an egocentric aspect to action generated by these be-
liefs; that is, the only thing that can act is “me.”
• Finally, intrusive experiences have, at some level, a salience, either an
irreducible epistemic affordance that cannot be dispelled or in terms of
aberrant precision; namely, the failure to suspend or attenuate attention
to certain kinds of cues. In OCD, for example, the inability to attenu-
ate arousal sensations manifests in the repetition of salient acts (such
as checking), despite the fact that these acts do not produce a lasting
reduction in uncertainty about the state of the world. (A computational
model of the neuronal underpinnings of recurrent intrusions in OCD is
provided in Appendix 13.1.)
The Insula and Functional Anatomy of Salience and Value
Now let us consider the above account from a systems neuroscience perspec-
tive. In this setting, salience can be thought of as an attribute of a cue (i.e.,
internal or external stimulus) deemed important to the individual in a given
context (Uddin 2014; Kahnt and Tobler 2017; Miyata 2019)—it is salient be-
cause of the potential for information gain and thus belief updating. Salience is
distinct from value in that the latter is a valenced or signed currency that varies
monotonically from negative to positive, whereas the former is an unsigned
currency (i.e., something is salient or not). This means that value and salience
are dissociable in terms of what they mean for behavior: both negative and
positive outcomes can be salient in the sense that experiences can change our
beliefs, even if they are unpleasant (Kahnt and Tobler 2017). As such, intrusive
experiences can be thought of as arising from an aberrant processing of inter-
nal and external stimuli with respect to the current (belief) state of the indi-
vidual. This salience misattribution leads to an overemphasis of one thought or
action over the current, ongoing cognitive process and subsequently influences
attentional capture, motivation, and goal-directed cognition. Importantly, the
unsigned nature of salience calculations necessitates that both appetitive and
aversive stimuli can sway the calculations of salience that ultimately influence
behavior.
There are many potential points at which biases can enter salience calcula-
tion. Ascribing salience to a given stimulus at a given time scale (Kennerley
et al. 2011) and within a given context (Heilbronner and Hayden 2016) re-
sults from integration across a wide range of processes, including attentional
(Menon and Uddin 2010), reward (Olney et al. 2018), affective (Etkin et al.
2011), and homeostatic regulation (Craig 2009).
Neurobiological instantiation of both ongoing and intrusive, highly salient
events occurs at many levels of the neuraxis. One highly interconnected hub that
seems to play a major role as an integrator or transmitter of the interoceptive and
exteroceptive environment is the insula. More specifically, the anterior insula

(and the von Economo neurons it contains) possesses the anatomical linkages
to support awareness and (together with its connections to, e.g., the anterior cin-
gulate; Craig 2009) the monitoring of the environment that is necessary (when
combined with the calculation of value) to assign behavioral relevance to the
event. In Tourette syndrome, for example, the insula may play a role in assign-
ing salience and aversiveness to premonitory urges (Conceição et al. 2017).
There are important, mostly bidirectional connections between the anterior
insula and key affective, cognitive, autonomic, and regulatory systems—com-
ponents which place the anterior insula in a unique position in the calculation
of salience (Critchley et al. 2005; Craig 2010; Nieuwenhuys 2012). In addition
to the posterior regions of the insula that receive predominantly somatosen-
sory inputs, homeostatic regulators enter via the hypothalamus and amygdala,
hedonic inputs from the nucleus accumbens and orbitofrontal cortex, and mo-
tivational, social, and cognitive information from anterior cingulate, ventro-
medial, and dorsolateral prefrontal cortex (Craig 2010). While anatomically
and functionally simplistic, this schema provides a framework uniquely plac-
ing the insula in the position to assess the relative weights of the environ-
mental processes in the assessment of attentional capture. The insula also has
important connections to motor regions that allow it to then drive behavior.
In fact, in Tourette syndrome, it may play a role in driving tics (Conceição et
al. 2017), whereas in addiction it may be driving craving (Naqvi and Bechara
2010; Naqvi et al. 2014).
Importantly, there appears to be a transdiagnostic component to the dys-
regulation of salience attribution (McTeague et al. 2016), as neuroimaging
studies have demonstrated anterior insula involvement across a number of di-
agnostic assignments in various disorders characterized by intrusive events,
including addiction (Naqvi et al. 2014), ADHD (Klein et al. 2013; Bubenzer-
Busch et al. 2016; Norman et al. 2016), autism (Gu et al. 2018), OCD (Zhu et
al. 2016), psychosis (Brosey and Woodward 2017), anxiety (Paulus and Stein
2006; Shiba et al. 2017), and depression (Ellard et al. 2018). This is potentially
important because it suggests that interoception plays a key role in all forms of
aberrant salience or precision attribution, as illustrated by the OCD example
above, and will be discussed in more detail below.
Interoceptive Contributions to Intrusive Experiences
Understanding intrusive experience at the level of brain systems will be incom-

plete without a consideration of the systems that underlie the self. Selfhood is
fundamental to the phenomenology of intrusive experiences (see Liu and Lau,
this volume). If intrusive experiences are to be understood as involuntary men-
tal phenomena that disrupt ongoing psychological narrative flow (see above),
one needs to have a sense of oneself as both an observer, experiencing such
intrusions, and as an agent perceiving the intrusions as unsolicited interrup-

tions of one’s sense of agency (e.g., I did not intend to have these thoughts,
or perform these actions, even though I recognize them as my thoughts, or
my actions). Importantly, the brain systems that regulate and represent bodily
physiology, or interoception, are considered to be at the core of selfhood.
Increasingly, there is appreciation that the self as a continuous coherent, uni-
tary representation is not the output of one specific single specialized system
within the brain. Experiences of selfhood are embodied and require coordina-
tion between dissociable brain systems, as revealed by careful experimentation
and in the symptomatic expression of particular psychiatric and neurologi-
cal “disorders of self,” such as depersonalization disorder (for a review, see
Fletcher and Fotopoulou 2015). Within a broad taxonomy, selfhood can be
parsed into minimal (embodied or biological) and extended (reflective and nar-
rative) components (akin to the first- and higher-order mechanisms described
by Liu and Lau, this volume). The sense of a core minimal self is proposed
to emerge from the integrative processing of sensory and motor signals from
the body. The frequent concomitant occurrence of sensory signals on the body
eventually gives rise to mental, predictive models of “owned” first-person
feelings of (bodily) sentience and presence (e.g., I exist and feel alive in this
body), agency (e.g., I was the author of this action), and ownership (e.g., this
bodily experience belongs to me) (Gallagher 2005; Seth et al. 2012). Extended
concepts of the self are built on embodied self-representation to encompass
the notion of the narrative or autobiographical self (Damasio 1999). Extended
selfhood affords the ability to make one’s self the object of explicit thoughts
irrespective of any particular experience or perspective in the here and now
(e.g., self-reference, I am a woman). More generally, this enables reflection
on one’s experiences across time, space, and person in counterfactual ways: I
have always been a woman, I anticipate being a woman tomorrow, I imagine
that I am a woman in the mind of others (Fotopoulou 2015). These notions rest
on the idea that while perception can be understood as the unconscious process
of hierarchical Bayesian inference on the (hidden) causes of sensory input,
more higher-order abilities for self (metacognition) or other mentalization or
reflection rest upon similar unconscious inferential processes of greater depth,
whereby the generative models refer not only to current sensory predictions
but also to predictions about the effects of actions, not yet executed, and bodily
or external situations, not yet encountered (e.g., Palmer et al. 2015).
Interoception is at the core of this hierarchical view of the self and, by exten-
sion, of psychopathological disorders of self-representation (for an overview,
see Khalsa et al. 2018). Indeed, there is growing theoretical acknowledgment
that core aspects of selfhood (Seth 2013; Fotopoulou and Tsakiris 2017) and
emotion (Gu et al. 2013; Seth 2013; Barrett et al. 2016) can be formalized
as the inferential processing of interoceptive signals, implemented within a
Bayesian/predictive coding framework (discussed above). Correspondingly,
models of how interoceptive inference is instantiated or regulated within the
brain can inform understanding of emotional and psychosomatic disorders,

such as anxiety, depression, and fatigue (Paulus and Stein 2006; Barrett et al.
2016; Stephan et al. 2016). Here we offer an overview of the neurocognitive
mechanisms by which interoceptive processes underpin self-representation,
psychopathology, and, in particular, intrusive experiences. We conclude with
an example of an eating disorder as an instance in which interoception influ-
ences the psychopathological expression of intrusive experiences within a hi-
erarchical predictive framework that encompasses self-conceptualization.
What Is Interoception?
Interoception encompasses afferent signaling, integrative processing, and cen-

tral representation of the internal physiological state of the body (Quadt et
al. 2018; for a discussion of alternative definitions, see Ceunen et al. 2016).
Interoception is the sensory component of homeostatic and allostatic control.
Homeostasis refers to the regulation of internal physiology, through which life
is sustained by maintenance of a more or less constant internal environment,
through supporting the dynamic metabolic needs of bodily tissues while ex-
cluding potential toxic or other threats to the integrity of the body (homeostatic
regulation; Cannon 1929). Cardiac output, blood oxygenation, hydration, tem-
perature regulation, and blood glucose are among the many parameters regu-
lated homeostatically. However, homeostatic interoceptive autonomic reflex
arcs alone are inefficient: better control of internal state is achieved through
allostasis, wherein the future state of the body is predicted and responses are
made in anticipation of future physiological states to mitigate unpredicted dys-
homeostatic states that threaten life (Sterling 2012).
Allostasis is informed by the integration of interoceptive information with
exteroceptive (about the external world) information for the predictive selec-
tion of autonomic/physiological and behavioral action or “policies,” which
ultimately ensure longer-term survival. For example, the set point of the ho-
meostatic baroreflex, which stabilizes blood perfusion of organs by regulat-
ing the heart’s beat-to-beat output, is allostatically adjusted to meet actual and
anticipated physical demands (e.g., if you see a snake or bear in the woods,
baroreflex suppression allows your heart rate and blood pressure to rise to-
gether to enhance skeletomuscular perfusion, facilitating the capacity for fight
and flight). From a more computational perspective, the most efficient way to
regulate homeostatic risk is to build a model of the body as separate from its
external environment, following the cybernetic idea that “every good regula-
tor of a system must be a model of that system” (Conant and Ashby 1970).
Ultimately, physiological control combines allostatic and homeostatic mech-
anisms, but both can be subsumed under homeostasis (Ramsay and Woods
2014). Allostatic anticipatory control requires an inferential model (hypotheses
about the causes of interoceptive inputs) of our own current and future (coun-
terfactual) bodily states in relation to states of the external world (including
other agents). The complexity is reduced by holding a set of prior “beliefs”

or more broadly generative models. Deviations from homeostatic ranges are
avoided by choosing in advance an appropriate sequence of actions (“poli-
cies”). These can be autonomic as well as behavioral and can cross different
systems. For example, you need to eat before you faint and you need to store
fat for future metabolic needs when resources may need to be allocated to other
tasks. These ideas are coherent with formal frameworks of brain function, such
as the Bayesian brain and active inference (discussed above). Details and dis-
cussion of the neural organization supporting interoceptive processing can be
found in Appendix 13.1.
Interoception and Intrusive Experiences
There are at least three ways in which interoception can impact upon intrusive
experiences:
1. It can provide context which can (a) have an impact on the permissive
threshold for the occurrence of intrusions (discussed above) and (b)
influence or constrain the content of what intrudes.
2. It can affect appraisal and control processes engaged by the intrusive
experience.
3. It can also act as content itself.
Moreover, these can interact to produce a self-sustained cycle of intrusive ex-
periences. In conceptualizing the impact of interoception on intrusive expe-
riences, it is helpful to conceptualize it within a hierarchical or dimensional
framework (see Table 13.1). Lowest in the hierarchy are the levels of physi-
ological arousal (indexed by heart rate, blood pressure, or electrodermal activ-
ity) and the bodily changes governed by homeostatic reflex arcs. These signal
the integrity and arousal state of the body through visceral afferent pathways.
Fluctuations in central signaling of bodily physiology (including both engage-
ment of ascending neuromodulatory systems and representation within pri-
mary “viscerosensory” insula, a cortical level) can thus provide the context
(Pt. 1 from the above list).
As a context, psychophysiological states (e.g., sickness, arousal, and alert-
ness) gate what enters the sensorium (Pt. 1a). For example, a heightened
state of cardiovascular arousal enhances the detection and appraisal of threat
(Garfinkel et al. 2014; Pezzulo et al. 2018) associated with symptoms of anxi-
ety; increased sympathetic electrodermal tone enhances occurrence of tics
in Tourette syndrome (Nagai et al. 2009). In addition, however, a particular
homeostatic context, such as hunger, can motivate relevant intrusions about
food (Pt. 1b; a specific example is given in the next section). Affective state
represents a more elaborated interoceptive context that can again change the
permissible threshold of intrusion.
Table 13.1 Dimensions of interoceptive measurement, adapted after Garfinkel et al. (2015). Psychological dimensions of interoception are given
in boldface; self-referential dimensions are capitalized.
Dimensional level Nature Index measures
EXECUTIVE Behavioral Shifting from interoceptive to exteroceptive attention (e.g., within dual
tasks or between tasks)
METACOGNITIVE Correspondence between subjective self- Receiver operating characteristic curves between task performance and
report and objective performance accuracy rated confidence
Correlational measures of task and confidence scores
Trait measures (e.g., correspondence between task performance and body
perception questionnaire score)
SENSIBILITY Subjective self-report Confidence measures on interoceptive tasks
Questionnaires probing interoceptive sensitivity
Accuracy Objective behavioral performance score Heartbeat detection tasks
Respiratory resistance load detection
Water load task
Balloon dilation of stomach/colon
Preconscious impact Behavioral, neural Cardiac modulation of eyeblink startle
on other processes Cardiac modulation of fear
Respiratory modulation of memory
Afferent signal Neural Visceral afferent nerve recording
Intracranial recording
Heartbeat evoked potential
Respiratory evoked potential
Neuroimaging
Bodily response Organ-level response Heart rate, heart rate variability, tachygastria, blood pressure, glucose, O2
and CO2 levels, etc.
Autonomic psychophysiology
265
Second, appraisal and control processes engaged by the intrusive experience

are impacted by higher-order cognitive levels of interoceptive representation,
likely supported within insula and cingulate cortices (Pt. 2). Higher-order cogni-
tive or psychological levels of interoception (highlighted in bold in Table 13.1)
refer to attention and appraisal directed at bodily processes themselves. These
encompass measures of interoceptive accuracy of objective (behavioral) sensi-
tivity to bodily responses, subjective (i.e., self-reported) interoceptive sensibility
to bodily signals, and metacognitive interoceptive insight (Garfinkel et al. 2015).
The latter two align with notions of expectation and interoceptive prediction er-
ror (“surprise”) and the precision weighting of interoceptive inputs, beliefs, and
policies. Interoceptive self-efficacy (Stephan et al. 2016) is a metacognitive rep-
resentation of self-efficacy.
Third, such a mental representation of bodily sensation may act as the con-
tent of the intrusion (Pt. 3). Salient bodily signals (e.g., breathlessness, heart
arrhythmia, urge to void, or visceral pain) necessarily attract attention and ap-
praisal. Upon appraisal, prior experience will determine if the intrusion per se
represents a major concern or acts as a driver for subsequent general persevera-
tive intrusions associated with overall health (e.g., health anxiety). Related to
this are the so-called quasi-interoceptive signals, such as rib pain (a somatic
sensation), which can be misinterpreted as a prelude to a heart attack, with anx-
iety again becoming amplified by the accompanying interoceptive sensations
of cardiorespiratory arousal as a consequence of the appraisal process (Clark
et al. 1997). Moreover, ephemeral interoceptive sensations can (through prior
associations) trigger emotional (e.g., panic or fear response PTSD) or drug-
related intrusive experiences such as craving (Goldstein et al. 2009; Garavan
2010). Similarly, the interoceptive feelings of premonitory urge, linked again
to representation within insular cortex, will trigger tics in Tourette syndrome
(Rae et al. 2018, 2019b).
Finally, an executive dimension of interoception contributes to intrusions
mostly through appraisal control processes (Pt. 2) which in turn can affect
the stickiness of the context (Pt. 1) and the interoceptive content (Pt. 3).
The executive dimension encompasses the capacity to shift between intero-
ceptive representations or away from interoceptive representations, aligned
with both precision weighting and policy selection. Such a capacity may
be evident in measures of lower levels of interoceptive signaling: for in-
stance, heart rate variability (a product of baroreflex regulation) is linked
to more general psychophysiological flexibility and is positively associated
with success in suppressing unwanted intrusive thoughts and memories, like
in PTSD (Gillie and Thayer 2014; Gillie et al. 2015). Conversely, intru-
sive perseverative cognition (worries and ruminations) and the capacity for
thought control are coupled to the inflexibility associated with blunted heart
rate variability, both in wakefulness and during sleep (Brosschot et al. 2010;
Meeten et al. 2016; Ottaviani et al. 2016; Ottaviani et al. 2017). It should
also be noted that in addition to all of these direct and indirect effects of
interoception on intrusive experiences, interoceptive signals may be evoked

under some circumstances as countermeasures to control intrusion, often
through physiological relaxation but sometimes using physiological arousal
(Nagai 2015).
Given this intimate relationship between interoception and intrusive experi-
ences, it is perhaps not surprising that disordered interoceptive processing is
reported across conditions associated with intrusive thinking. In anxiety dis-
orders, increasing evidence indicates an association between anxiety symp-
toms and a mismatch between subjective (sensibility) and objective (accuracy)
measures of cardiac interoception—a metacognitive interoceptive deficit (trait
interoceptive prediction error) (Garfinkel et al. 2015) that is also relevant to
symptoms in Tourette syndrome (Rae et al. 2019b), autism (Garfinkel et al.
2016), and, if extended to measures of choice, addiction (Moeller et al. 2014).
Moreover, intrusive dissociative experiences, consistent with a fundamental
self-disturbance in self-representation, are associated with lower-level intero-
ceptive abnormalities (Schulz et al. 2016). Below we present an example of
how abnormalities in interoception can act as the content and character of an
intrusive experience.
Intrusive Experiences in an Ego-Syntonic Disorder

Exemplified by Anorexia Nervosa
Patients with anorexia nervosa report thoughts, bodily experiences, and mental
images that they consider as involuntary and intrusive to other goals, even
though these may not always be unpleasant in themselves and may, in fact,
constitute most people’s everyday experiences. For example, a patient de-
scribed the feeling of a full stomach as intruding on her mental concentration
(Skårderud 2007:127):
Some days ago, I should have had a meeting with my boss. I was anxious about
this. Then I decided to vomit. I couldn’t stand having the lunch in my stomach. I
cannot have anything in my stomach, because then I cannot concentrate. I need
to be empty to feel alert.
Similar experiences of hunger or satiation and other interoceptive sensations

are frequently experienced as intrusive by individuals with anorexia, while
their attempts to control their eating and body weight and to “silence” any
relevant bodily needs are seen as compatible with the goal of building a co-
herent and stable self. This treatment-resistant concordance in eating disor-
ders between symptoms and a sense of self is referred to as ego-syntonicity
(Gregertsen et al. 2017). Unlike in (ego-dystonic) disorders like OCD, where
symptoms are seen as intruding into one’s other everyday goals, anorexia ex-
emplifies a psychiatric disorder where symptoms are not viewed as intrusions
into one’s life; instead, necessary bodily functions, and particularly interocep-
tive experiences, are experienced as intrusive.
Drawing on the framework outlined earlier in the chapter as well as from

knowledge of the brain systems that support homeostatic and allostatic con-
trol (outlined above), intrusive experiences in eating disorders, particularly in
anorexia nervosa, can be understood as a failure to model interoceptive states
at a homeostatic level, due to a deeper failure in the regulation of metabolism
(notably adiposity or fat storage) at an allostatic level (i.e., a failure to opti-
mize flexibly the precision weighting of allostatic control policies). In other
words, patients with anorexia nervosa may not be able to correctly predict
and regulate adiposity (and metabolism more generally), leading to a chronic
dyshomeostatic state that evokes aberrant metacognitive beliefs about the low
efficacy of their autobiographical self: “I cannot eat now because I will then
lose control over my eating and store excessive fat” (see Figure 13.6). Recent
converging evidence highlights wide dysregulation across neuromodulatory
systems in eating disorders, including hormones and neuropeptides involved in
the regulation of metabolic states (see Figure 13.6; Gorwood et al. 2016), and a
large-scale genetic study implicating metabolic (alongside psychiatric) factors
in pathoetiology of anorexia nervosa (Watson et al. 2019). Neurocomputational
formulations of allostasis, that is, predictive, counterfactual interoceptive con-
trol (Stephan et al. 2016), suggest that allostasis requires a temporary change
or suspension of homeostatic set points, effectively altering the priors (beliefs)
of the relevant homeostatic reflex arc (e.g., the expectation of a meal will drop
Innate and Developmental Antecedents

Precision optimization difficulties
(candidates: serotonin, oxytocin, testosterone, estrogen, ghrelin, or leptin abnormalities)
Allostatic/Metacognitive Models about Metabolic Regulation

AIC, ACC, OFC
= I am going to lose control of my weight if I store any fat
Metabolic states of the body change Experienced as intrusive

in time anyway Reinforcing beliefs
Interoception =
Hypothalamus Ongoing PEs
Brain stem Homeostatic predictive models
about hunger/satiation
PIC and MIC
Action = = Dangerous, must be kept to a
Restriction minimum
States the brain is trying to predict
Eating restriction
Figure 13.6 Schematic depiction of a predictive coding account of intrusive intero-

ceptive experiences in anorexia nervosa: AIC (anterior insular cortex), ACC (anterior
cingulate cortex), OFC (orbitofrontal cortex), PIC (posterior insulate cortex), MIC (me-
dial insular cortex), PE (prediction error).
blood glucose levels to mitigate the hyperglycemia that follows eating). In

the brain, allostatic coupling of behavioral policy with internal physiology is
supported by regions including the anterior insula and dorsal and subgenual
anterior cingulate cortices. These manifest three properties (Stephan et al.
2016): (a) access to estimates of bodily state (interoception), (b) the capacity
to generate predictions over longer time scales, and (c) anatomical connec-
tions (descending visceromotor outputs) that can convey sustained changes in
homeostatic beliefs instantiated by more reactive humoral/autonomic reflex
response arcs within hypothalamus, brain stem, and periphery. Thus, func-
tional abnormalities within these regions may lead to inappropriate adjust-
ments to specific physiological parameters (e.g., glucose levels before or after
meals), leading to persistent prediction errors driving abnormal eating habits.
For anorexia nervosa, eating control will always be suboptimal for regulating
metabolic and interoceptive states since these necessarily fluctuate in time.
Persistent exacerbated interoceptive feelings of hunger and satiation are expe-
rienced as ongoing intrusive experiences that interfere with the ego-syntonic
goal of a rigid control of body fat, achieved by eating restraint, exercise, and/
or vomiting. These acts in themselves and their interoceptive consequences
reinforce the homeostatic beliefs of patients regarding the unpredictable and
intrusive nature of hunger and satiation signals.
Several studies have indeed shown abnormalities in correctly predicting
and experiencing interoceptive states in anorexia nervosa, including, for ex-
ample, cardiac signals, satiation and affective touch, and the related brain
function abnormalities best tracked by the anterior insular cortex and related
limbic and prefrontal areas (Crucianelli et al. 2016; Bischoff-Grethe et al.
2018; Khalsa et al. 2018). Such abnormalities have been linked to persistent
prediction errors about interoception and a dysregulated ability to adequately
sense what is happening in the body resulting in a turbulent reference state;
that is, a “noisy baseline” (Paulus and Stein 2010). This may explain why
patients experience all those states as intrusive experiences of the body that
need to be controlled by eating restriction, exercise, or vomiting (see Figure
13.6). These attempts to actively restrict and control hunger and satiation in
turn lead to starvation and further maintenance mechanisms (starvation damp-
ens hunger and slows down cognitive processing along with further com-
plications). According to the above speculations, a fundamental difficulty in
reducing interoceptive uncertainty via allostatic control would be at the heart
of why otherwise normal feelings of hunger or satiation are experienced as
intrusive and as “out of control.”
Relevance of Stability and Flexibility to Intrusion Experiences
Balancing stability and flexibility in the brain is critical for individuals to max-
imize exploitation and exploration of their environment. Working memory and
associative learning models provide a psychological and neural framework in

which the concepts of flexibility and stability can be understood (Hochreiter
and Schmidhuber 1997; Frank et al. 2001; Oberauer 2013). Biophysically de-
tailed computational models have also investigated how dynamical interactions
between different neuronal populations in cortex may promote stability ver-
sus flexibility (Durstewitz et al. 2000; Wang 2001). In addition, a large body
of empirical evidence has implicated specific neural structures and neuro-
modulators in behavioral flexibility (Robbins 2005; Cools and D’Esposito
2011). In particular, a central role is played by the prefrontal cortex and its
interactions with the rest of the brain, especially the specialized processing
modules of the posterior cortex, including parietal (spatial attention) and in-
ferotemporal (feature attention) areas; the declarative memory systems in the
temporal lobes, including the rhinal cortex (recognition memory) and hippo-
campus (scene/episodic memory); and the language processing modules such
as Wernicke’s area, specialized in the comprehension of speech, and Broca’s
speech and production area. In addition, the prefrontal cortex interacts with
subcortical structures such as the limbic structures involved in the processing
of motivational and emotional cues as well as the orchestration of behavioral,
autonomic, and endocrine responses, including the amygdala, hypothalamus,
and brain stem centers; the basal ganglia, which are involved in the higher-
order control of thought (see below) and action; and the neuromodulatory sys-
tems of the reticular core of the brain, including monoamine and cholinergic
cell groups in the midbrain and hindbrain.
Working Memory Models
As an example of how these interactions could support a balance between

stability and flexibility, let us consider the case of working memory. In mod-
els of working memory, stability (i.e., stable goal-oriented performance)
can be maintained by holding temporally stable representations of our
goals. Goals for action can reside at different levels of task abstraction and
unfold over different timescales. Importantly, however, a goal held in work-
ing memory can include the goal of meeting the requirement of specific
tasks. As such, our ability to hold this goal in memory, available for use as
a control signal, allows for stable task performance. Likewise, our ability to
update working memory (i.e., to shift goals as context demands) is impor-
tant for flexibility.
The control of working memory is often conceptualized as a gate that is
distinct from the memory store itself (Hochreiter and Schmidhuber 1997).
Closing an input gate against distracting information prevents its access to
working memory, keeping the current contents available as control signals;
this gating function promotes stability. In contrast, opening the gate enables
the updating of working memory and allows new contextual information to
modify behavior; this gating function promotes flexibility.
Though different mechanisms for working memory gating have been pro-
posed (e.g., Wang et al. 2004; Zhu et al. 2018), one influential model has
focused on frontostriatal interactions (Frank et al. 2001). This circuit is sche-
matized in Figure 13.7 (O’Reilly 2006). The corticostriatal model of working
memory gating proposes that the prefrontal cortex supports information main-
tenance, whereas the striatum-pallidal-thalamic pathway implements gating by
regulating what information is allowed in and out of working memory.
Based on this model, spontaneous, unwanted events could be experienced
as intrusions when the gate to working memory is breached and the intru-
sion supplants ongoing working memory processes. Once this occurs, intru-
sive events serve as signals to drive other cognitive processes and actions.
Thus, the integrity of the working memory gating is paramount for mitigating
against intrusive experiences. For example, by preventing an unwanted expe-
rience from updating to working memory or by inhibiting their influence on
output control signals, one could stop the negative cycle of behaviors that can
result from intrusive experiences. These gating mechanisms could be global
(like the fast, inhibitory mechanisms supported by the hyperdirect pathway
that can affect multiple processes simultaneously) or selective, supported by
both the direct and indirect pathways, schematized in Figure 13.7 as the Go
and No-Go pathways. Coordination among multiple corticostriatal loops can
also be a mechanism for working memory operations in separate prefrontal
areas to carry out complex, sequential, and hierarchically structured tasks (for
a review, see Badre and Nee 2018).
(a) Frontal cortex maintains (b) Frontal cortex working

information memory gets updated
Posterior Frontal Posterior Frontal
cortex cortex cortex cortex
Not active Active
Striatum Striatum
Thalamus Thalamus
Go No-Go VA, VL, MD Go No-Go VA, VL, MD
Gate closed GPe Gate open GPe

(No-Go) (Go)
SNr SNr
Excitatory Inhibitory
Figure 13.7 Schematic depicting a mechanism of working memory gating through

corticostriatal interactions. Inhibition (a) or disinhibition (b) of thalamocortical dynam-
ics through the striatum can regulate gate closing and opening, respectively: VA (ven-
tral area), VL (ventrolateral), MD (medial dorsal), GPe (globus pallidus external), SNr
(substantia nigra pars reticulata). Reprinted with permission from O’Reilly (2006).
The type of gating one selects can be thought of as a gating policy, in the
same sense as defined above. As such, the match of the right gating policy
to the particular dynamics of the situation is a key determinant of success-
ful control (Bhandari and Badre 2018). For example, when confronted with
an unwanted memory, one could deploy a global suppression to prevent it
from entering working memory or instead attempt to selectively input another
thought into working memory in its place. The consequences of these policies
on memory or the ongoing impacts of the triggering event (both in this instance
or in the future) might differ depending on the gating strategy that is selected.
Thus, pathologies could arise as a result of any of the following:
• Items seeking to enter working memory are sufficiently salient or val-
ued and will therefore breach the gating mechanism to update working
memory (breach intrusion).
• Gating itself is weak and thus items access working memory, even if
they are not adaptive or helpful to the individual (permissive intrusion).
• Mechanisms involved in maintaining stability (other than gating) are
too strong and thus do not allow working memory to be updated once
an intrusive experience has occurred.
• The wrong gating policy is selected given the nature or dynamics of
the intrusion.
Associative Learning Models
Corticostriatal circuits are also central to associative learning models

(Balleine and Dickinson 1998). Two control processes have been identi-
fied that are engaged in the control of goal-directed and habitual actions
and which are mediated by distinct parallel circuits through the basal gan-
glia; in some circumstances, they compete with one another (Balleine et
al. 2009). The goal-directed network is engaged rapidly with changes in
the environment, incorporates the cortical working memory process de-
scribed above, and utilizes this network to encode the action–outcome
associations that mediate goal-directed action in a region of dorsomedial
striatum. Generally speaking, this network relies on this prefrontal-dor-
somedial-striatal (or caudate) pathway and feedback to the cortex via the
substantia nigra pars reticulata and mediodorsal thalamus (Balleine and
O’Doherty 2010) to encode and utilize novel solutions to problems pre-
sented by a changing environment. It also functions to inhibit older, more
routine and outdated solutions, particularly the performance of habitual
actions centered on the sensorimotor cortices and putamen or dorsolateral
striatum (Graybiel 2008), when these have or are likely to have aversive
consequences. If the goal-directed circuit is altered (e.g., through dam-
age, disease, or drugs), inhibition can be reduced or mistimed, resulting in
dysregulation of habits (even in the presence of aversive consequences),
and producing intrusive experiences. An increased reliance on habits may

not only apply to the compulsive acts seen in OCD (Saxena et al. 1998;
Robbins et al. 2019), but also the persistent motor habits (tics) associ-
ated with Tourette syndrome (Maia and Conceição 2017, 2018) as well as
craving and compulsive drug use in addiction (Everitt and Robbins 2016;
Furlong et al. 2017).
There are, however, other important features that are controlled by the
goal-directed circuit, particularly by the dorsomedial striatal component of
that circuit. As mentioned, considerable evidence suggests that the prefrontal
working memory systems provide inputs to the striatum that mediate the
plasticity necessary to encode goal-directed actions in the posterior segment
of the dorsomedial striatum (reviewed in Balleine and O’Doherty 2010).
However, to allow this large structure to encode more than one action–out-
come association, plasticity associated with new action–outcome learning
needs to be segregated from prior learning. It appears that this segregation
is achieved via state-related information provided by inputs to the striatum
from the parafascicular thalamus (Bradfield et al. 2013). This input onto the
tonically active striatal cholinergic interneurons causes them to pause, allow-
ing the principal neurons (the spiny projection neurons) relief from inhibi-
tion induced by tonic acetylcholine release. During this pause, cortical and
midbrain dopaminergic inputs to the dorsomedial striatum can combine to
induce plasticity in the spiny projection neurons. Accordingly, changes in
action–outcome contingency provoke changes in the patterned input from
the parafascicular thalamus, leading to plasticity changes in the targeted dor-
somedial-striatal region.
Importantly, evidence suggests that the retrieval of specific action–outcome
ensembles for performance is mediated by state-related information, based
largely on outcome-related information (Bradfield et al. 2015) conveyed to
the striatum, not by the parafascicular thalamus but via inputs from the orbito-
frontal cortices (Gremel and Costa 2013; Bradfield et al. 2015; Stalnaker et al.
2016). Thus, accurate retrieval of specific action–outcome associations will be
determined by the fidelity of this orbitofrontal cortical input: as a consequence,
changes in orbitofrontal cortex activity (e.g., in OCD) could result in faulty
retrieval, causing changes in flexibility (described above) and leading to the
intrusion of unwanted information. Retrieval can become “frozen” if the orbi-
tofrontal cortex gets “stuck” in a given state (see Appendix 13.1); alternatively,
it could become highly, temporally disparate if activity in the orbitofrontal
cortex fluctuates rapidly and unpredictably.
This type of state information features heavily in computational accounts,
particularly model-based reinforcement learning accounts of goal-directed ac-
tion. Such accounts provide information about state transitions for retrieval and
could be seen as the computational implementation of these ideas (Wilson et
al. 2014). See Appendix 13.1 for further computational modeling of recurrent
intrusions in OCD focusing on neuromodulation within orbitofrontal cortex.
Summary
Translating findings across different levels of analysis, including computa-

tional, psychological, neurobiological, and physiological, is challenging.
However, to understand the nature of intrusive experiences and to develop
effective treatments, such translation is essential. A first step in this transla-
tion is to use a common language. To this end, we have attempted to define
all terms and concepts, especially where multiple, related, but somewhat dis-
tinct meanings exist. Salience is one such example of a term that is often
used broadly to refer to the quality of being particularly noticeable, but in a
Bayesian framework is specifically used to refer to the value afforded to un-
certainty resolution. The models we discuss provide explanations for a range
of intrusive experiences: from the obsessions and compulsions of OCD and
drug and emotion-related intrusions in addiction and PTSD to intrusions of
thoughts, bodily experiences, and mental images in anorexia nervosa. We
focused on two major networks, frontostriatal and insula-cingulate, to illus-
trate how imbalances in these networks can lead to intrusive experiences.
Whenever possible, overlap between different models and levels of analysis
have been highlighted to provide a systems overview of how intrusive experi-
ences across a range of distinct psychiatric, neurodevelopmental, and neu-
rological disorders may emerge as a consequence of dysfunction at different
levels of the nervous system.
Appendix 13.1
Active Inference
This appendix provides a technical description of belief updating under ac-

tive inference. One useful aspect of treating “trains of thought” as “planning”
under a generative model is that one can always express a generative model
as a graphical model (Figure 13.A1). This is important because a graphical
model can be used to understand the computational architecture of neuro-
nal message passing in the brain. For every graphical model that specifies
the states and outcomes in play and their conditional dependencies, there is
an associated factor graph that provides, and must be supported by, unam-
biguous specifications of the architecture (e.g., neuronal connectivity) and
message passing (e.g., neurophysiology); for details, see Figure 13.A2 and
Friston et al. (2017).
In brief, the sorts of generative models commonly used to explain plan-
ning as inference are usually based on partially observed Markov decision pro-
cess models. Crucially, in these generative models, discrete states of the world
evolve over time in a way that depends upon action.
Expected Free Energy, Salience, and Value
For technical readers, expected free energy can be decomposed into an epis-
temic, information-seeking, uncertainty-reducing part (intrinsic value) and a
pragmatic, goal-seeking, instrumental part (extrinsic value). Formally, the ex-
pected free energy for a particular policy π at time τ in the future can be ex-
pressed as in terms of beliefs Q s , o about future states sτ and outcomes oτ:
G , E ln Q s | o , ln Q s | E ln P (o ) . (13.A1)
intrinsic value extrinsic value
Extrinsic (instrumental) value is simply the expected value of a policy defined

in terms of outcomes that are preferred a priori. The more interesting part is
the uncertainty-resolving or intrinsic (epistemic) value, variously referred to
as relative entropy, mutual information, information gain, Bayesian surprise,
intrinsic motivation, or value of information expected under a particular policy
(Barlow 1961; Howard 1966; Optican and Richmond 1987; Linsker 1990; Itti
and Baldi 2009).
Intrinsic (epistemic) value can be regarded as salience. Formally, this
means that salience is the Kullback-Leibler (KL) divergence between pos-
terior beliefs about hidden states with and without observations solicited by
a particular act (or policy). The reason this divergence is associated with sa-
lience stems from the visual neurosciences, where the salience of a potential
location for a saccadic fixation is known as Bayesian surprise (Itti and Baldi
2009; Sun et al. 2011; Barto et al. 2013). In robotics and machine learn-
ing, the information gain or Bayesian surprise is known as intrinsic motiva-
tion or value (Ryan and Deci 1985; Eccles and Wigfield 2002; Oudeyer and
Kaplan 2007; Schmidhuber 2010; Barto et al. 2013). It is also referred to as
epistemic value or epistemic affordance (Parr and Friston 2017). Epistemic
affordance appeals to Gibsonian notions of affordance: it is the resolution of
uncertainty afforded by a particular act: “What would I learn by looking over
there?” On a psychological interpretation, intrinsic value can also be associ-
ated with incentive salience (Berridge and Robinson 1998; McClure et al.
2003). Exactly the same kind of mathematical arguments can be applied not
just to beliefs about states in the world but also the parameters of the genera-
tive model. These parameters encode contingencies and laws governing the
evolution of states or their mapping to observations. In this setting, salience
becomes novelty; namely, the information gain afforded by knowing “what
would happen if I did that?”
The factor graph in Figure 13.A2 is used to pass messages among the nodes
(e.g., neuronal populations) to minimize free energy per se; in other words,
to maximize the evidence for any given generative model of how outcomes
were generated. This leads to biologically plausible message-passing schemes
of the sort studied in terms of evidence accumulation and predictive coding
G
276
Generative model
P o1:W , s1:W , S P ( s1 ) P (S )W P(oW | sW ) P( sW | sW 1 , S )
S
1 P oW | sW Cat ( A)
2 P sW 1 | sW , S Cat (BS ,W )
3 P s1 Cat (D) Factors D s1 BS ,1 s2 BS ,2 s3
4 P S V (G ) (likelihood and empirical priors)
A A A
Q sW | S Cat (sS ,W )
Q S Cat ( ʌ) Approximate posterior o1 o2 o3
ʌ
1
2 3 4
sS ,W V (ln BS ,W 1sS ,W 1 ln BS ,W sS ,W 1 ln A oW ) G
4
1 ʌ V (G ) 3 5 2 5 2 5
sS ,1 sS ,1 sS ,2 sS ,2 sS ,3 sS ,3
D 2 = 3 BS ,1 2 = 3 BS ,2 2 =
5 GS , ,
¦W oS W (ln oS W CW H sS W ) , sS ,1 sS ,2 sS ,3
oS ,W AsS ,W Belief updating 4 4 4

A A A
1
uW max u ʌ [U S ,W u]
Action selection o1 o2 o3
Figure 13.A1 A generative model for discrete states and outcomes. Upper left panel: these equations specify a generative model. A generative
model is the joint probability, P, of outcomes or consequences and their (latent or hidden) causes, see first equation. Usually, the model is expressed
in terms of a likelihood (the probability of consequences given causes) and priors over causes. When a prior depends upon a random variable it is
called an empirical prior. Here, the likelihood is specified by a matrix A, whose elements are the probability of an outcome under every combina-
tion of hidden states. Cat denotes a categorical probability distribution. The empirical priors pertain to probabilistic transitions (in the B matrix)
among hidden states that can depend upon actions, which are determined by policies (sequences of actions encoded by π). The key aspect of this
generative model is that policies are more probable a priori if they minimize the (time integral of) expected free energy G, which depends on
prior preferences about outcomes or costs encoded in C and the uncertainty or ambiguity about outcomes under each state, encoded by H. Finally,
the vector D specifies the initial state. This completes the specification of the model in terms of model parameters that constitute A, B, C, and D.
Bayesian model inversion refers to the inverse mapping from consequences to causes (i.e., estimating the hidden states and other variables that
cause outcomes). In approximate Bayesian inference, one specifies the form of an approximate posterior distribution Q. This particular form in
this figure uses a mean field approximation, in which posterior beliefs are approximated by the product of marginal distributions over time points.
Subscripts index time (or policy), italic variables represent hidden states, and bold variables indicate expectations about those states. Upper right
panel: this Bayesian network or graphical model represents the conditional dependencies among hidden states and how they cause outcomes. Open
circles are random variables (hidden states and policies), filled circles denote observable outcomes, and squares indicate fixed or known variables,
such as the model parameters. Lower left panel: these equalities are the belief updates mediating approximate Bayesian inference and action selec-
tion. Lower right panel: this is an equivalent representation of the Bayesian network in terms of a Forney or normal style factor graph. Here the
nodes (square boxes) correspond to factors and the edges are associated with unknown variables. Filled squares denote observable outcomes. The
edges are labeled in terms of the sufficient statistics of their marginal posteriors (see approximate posterior). Factors have been labeled intuitively
in terms of the parameters encoding the associated probability distributions (on the upper left). The circled numbers correspond to the messages
that are passed from nodes to edges (the labels are placed on the edge that carries the message from each node). These correspond to the messages
implicit in the belief updates (lower left). This figure is based on Friston et al. (2017).
277
G
278
u1 u2
D s1 Bu1 s2 B u2 s3
A A A
sS ,W V (ln BS ,W 1sS ,W 1 ln BS ,W sS ,W 1 ln A oW )
ʌ V (G )
o1 o2 o3
GS , ,
¦W oS W (ln oS W CW ) H sS W ,
oS ,W AsS ,W Belief updating A A A
sS ,1 sS ,2 sS ,3
uW max u ʌ [U S ,W u]
Action selection BS ,2
D BS ,1
= = =
Figure 13.A2 The generative process and model. This figure reproduces the Bayesian network and Forney factor graph of Figure 13.A1; how-
ever, here the Bayesian network describes the process that generates data, as opposed to the generative model of data. This means that we can link
the two graphs to show how the policy half-edge of Figure 13.A1 couples back to the generative process (by generating an action that determines
state transitions). The selected action corresponds to the most probable action under posterior beliefs about action sequences or policies. Here,
the message labels have been replaced with little arrows to emphasize the circular causality implicit in active inference: the real world (red box)
generates a sequence of outcomes that induce message passing and belief propagation to inform (approximate) posterior beliefs about policies (that
also depend upon prior preferences and epistemic value). These policies then determine action, which generates new outcomes as time progresses,
thereby closing the action perception cycle. This figure is based on Friston et al. (2017).
279
(Srinivasan et al. 1982; Rao and Ballard 1999; Huk and Shadlen 2005; Beck et
al. 2008; Bastos et al. 2012; Egner and Summerfield 2013; de Lafuente et al.
2015; Kira et al. 2015; Shipp 2016). In terms of the parameters of the genera-
tive model, associative plasticity is the corresponding belief update for neuro-
nal connections (Friston et al. 2016).
Precision and Parameters
Of particular interest here are the parameters that link states to outcomes and
states at one point in time to states at the next point in time. In Figure 13.A1,
these are simply matrices of probabilities encoding the likelihood mapping
from states to outcomes A and the transition probabilities to one state to the
next B (which depend upon a particular policy).
Neurobiologically, these matrices play the role of connectivity matrices,
which play an important role in sensory data assimilation and subsequent
planning based on beliefs about the consequences of any action. Furthermore,
each column of these matrices has a precision. Precision, in this instance, re-
flects the fidelity or confidence about the outcome (or subsequent state) given
the current state of the world. A very precise mapping means that we can be
almost 100% confident that this will happen given that state, while a very
imprecise mapping means that all outcomes (all subsequent states) are equally
likely. For discrete space models, one can express the likelihood and priors in
terms of inverse temperature or softmax parameters with the following form,
where ( ) is a softmax function or normalized exponential:
P o |s o ln A
P s | s 1, s ln B (13.A2)
P G .
Neural Organization of Interoceptive Processing
The control and representation of internal bodily physiology is instantiated

throughout the neuraxis (for reviews, see Craig 2003; Critchley and Harrison
2013). While ganglionic and spinal reflexes support proximate physiologi-
cal regulation, the brain orchestrates homeostatic control and allostatic re-
sponses across bodily organs, integrating control with behavioral demand.
The brain receives interoceptive information about the internal state of the
body via neural afferent and humoral interoceptive routes (for details, see
Figure 13.A3). Somatosensory pathways also contribute to quasi-interocep-
tive sensation of bodily physiology (e.g., via heart beating against the chest
wall, rib motion, pharyngeal airflow) and to referred pain (e.g., angina felt in
shoulder). Chemosensory signaling (including O2/CO2, hormones, cytokines,
blood pH, glucose, and hydration) occurs through central blood sampling at
Ascending neuromodulatory projections

Anterior cingulate
INSULA
Midbrain Orbital and vmPFC

Amygdala
Hypothalamus
Pons
Medulla oblongata Glossopharyngeal XI

including NTS
Carotid baroreceptors
Vagus X Aortic baroreceptors
Lamina 1 Sinoatrial node
spino- atrioventricular node
thalamic
tract Myocardium
Lungs and diaphragm
Stomach
Small bowel
Colon
Adrenals
Bladder
Genitals
Figure 13.A3 Schematic illustration of feed-forward neural interoceptive pathways.

Peripheral afferents in cranial nerves X (vagus) and X1 (glossopharyngeal) and those
following sympathetic nerves to spine (ascending laminar1) converge in the medullary
nucleus of the solitary tract (NTS). Here, local connections support homeostatic reflexes
(e.g., baroreflex) modulating autonomic outflow. Interoceptive information passes via
a primary thalamocortical route to insula (shown in gray; viscerosensory cortex) where
integrative processing builds a representation within anterior insular that is consciously
accessible and can give rise to potentially intrusive interoceptive and affective feel-
ings. Secondary interoceptive channels include (1) a subcortical route to hypothalamic
and basal ganglia (including amygdala), modulating ascending widespread monoamine
projections from midbrain (shown in light green) and (2) a thalamocortical route to
visceromotor anterior cingulate cortex. Connections to orbitofrontal and ventromedial
prefrontal cortices (vmPFC) offer another putative source of intrusive motivated feel-
ings related to selection of action policies.
paraventricular organs and hypothalamus and may engender powerful motiva-

tional and arousal states (e.g., air hunger) with correspondingly intense feel-
ings. The interoceptive representation within insular cortex shows a partial
viscerotopy and connections follow a posterior-anterior and dorsal-ventral pro-
gression with increasing opportunity for cross-modal integration (Craig 2003,
2009; Evrard 2019). Anterior insula is most implicated in supporting conscious
access to interoceptive sensations and associated emotional and motivational
feelings (Critchley et al. 2004), including the urge-to-tic in Tourette syndrome

(Conceição et al. 2017) and drug cravings (Goldstein et al. 2009; Garavan
2010). Reciprocal connections between anterior insula and “visceromotor”
rostral cingulate regions (both “allostatic” dorsal anterior cingulate and “ho-
meostatic” subgenual cingulate) represent a putative functional architecture
for higher-order predictive regulation of bodily states (Critchley et al. 2004;
Critchley et al. 2005; Medford and Critchley 2010). As described above, ante-
rior insula and dorsal anterior cingulate are key hubs within the so-called sa-
lience network (highlighting the motivational primacy of interoception, where
salience is the epistemic value afforded by uncertainty resolution of Bayesian
surprise; see above and Fedota and Stein, this volume).
Example of a Computational Model to Explain Recurrent Intrusions

in Obsessive-Compulsive Disorder
In addition to being intrusive, obsessions in OCD are both recurrent and

“sticky” in the sense that they are difficult to shake from mind. From a dynami-
cal systems perspective, these characteristics seem to suggest that obsessions
correspond to attractors (Rolls et al. 2008; Maia and McClelland 2012; Rolls
2012; Maia and Cano-Colino 2015); that is, states toward which a system tends
and from which it may have difficulty escaping.
OCD prominently involves disturbances in the orbitofrontal cortex and con-
nected regions (Maia et al. 2008). Neurochemically, OCD may be associated
with low serotonin and/or high glutamate. A biophysically detailed computa-
tional model of serotonin and glutamate modulation of the orbitofrontal cortex
showed that both low serotonin and high glutamate tend to create excessively
strong attractors in the orbitofrontal cortex (Figure 13.A4). The network tends
to fall into these attractors and then has difficulty escaping from them. This is
consistent with the perseverative responding to a previously rewarded visual
stimulus displayed by marmoset monkeys following depletions of serotonin in
the orbitofrontal cortex, either following reversal (Clarke et al. 2006) or extinc-
tion (Walker et al. 2009) of the association between the stimulus and reward.
In these simulations, neuronal activity was elicited by “manually” activat-
ing subsets of neurons. A more complete design would also have to incorporate
the endogenous gating of information into (and out of) this local network, as
was described above in the context of working memory. In addition, there are
complex interactions between neuromodulatory levels and their effects across
interacting brain structures. For example, the extent to which a monkey dis-
plays perseverative responding depends not only on low levels of serotonin
in the orbitofrontal cortex but also on high levels of dopamine in the striatum
(Groman et al. 2013). Moreover, alterations in these neuromodulators at the
level of the orbitofrontal cortex can have profound opposing influences on
the levels of the same or different neuromodulators in other structures, includ-
ing the striatum (Roberts et al. 1994; Clarke et al. 2014) and the amygdala
(a) Neuron
Neuron
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Time (s) Time (s) Time (s) Time (s) Time (s)
(b) Normal conditions Lower 5-HT level in all neurons
20 20 5-HT Ļ 35 5-HT Ļ

30
Firing rate
Firing rate
Firing rate
10 10 20
10
0 0 0
1000 1000 1000
60 70 60 70 40 50 60
70
Neu500 30 40 50 Neu500 30 40 50 Neu500 0 0 10 20 30 rial
ron 0 0 10 20 Trial ron 0 0 10 20 Trial ron T
(c)
Switch
Perseveration
Neuron
Memory loss
1
0.8
Proportion of traits
Neuron
0.6
0.4
Neuron
0.2
0
0 1 2 3 4 5 –30 –25 –20 –15 –10 –5 0
Time (s) &KDQJHUHODWLYHWRSK\VLRORJLFDO+7
Figure 13.A4 A computational model of the role of serotonin (5-HT) in the orbito-
frontal cortex in OCD, adapted after Maia and Cano-Colino (2015). (a) Illustration of
the process of entrenchment of patterns of neuronal activity, taken to correspond to
obsessions or, in less pathological cases, “habits of thought.” Each plot represents a
population of neurons; the dots along each line represent the action potentials for one
neuron. The network stochastically develops patterns of activity (“bumps”). Each bump
elicits strengthening of the synapses between the neurons that were active in that bump
through Hebbian learning, thereby developing attractors (orange bands). The more fre-
quently a bump occurs, the more likely it is that it will reoccur (see last three plots).
(b) Effects of reducing serotonin on the tendency to develop and fall into excessively
strong attractors. Under normal circumstances, the network develops bumps at varying
places over time (left panel). Under low levels of serotonin, however, the network tends
to develop excessively strong attractors into which it repeatedly falls (middle panel).
Moreover, there is a dose-response effect, such that reducing serotonin further causes
even stronger attractors to develop (right panel). Increasing glutamate has the same
effect as decreasing serotonin (not shown). (c) Low levels of serotonin cause the at-
tractors to become excessively stable. Simulated activation of a set of neurons elicited
a bump, followed by activation of a different set of neurons. Under normal conditions,
the network’s pattern of activity flexibly shifts to the state represented by the new bump
(blue). Under low levels of serotonin, the network fails to shift to the new bump, result-
ing in perseverative activation of the prior bump (brown). (continued on next page)
Figure 13.A4 (continued) Importantly, low levels of serotonin increase such perse-
verative errors (brown) without affecting a different type of error in which the network
simply loses the memory of what it was initially representing (green). The latter error,
which is more reminiscent of disorders in which there is difficulty in keeping items in
working memory (e.g., ADHD), is not affected by the serotonin manipulations.
(Roberts and colleagues, unpublished), which may exacerbate the possibility

of intrusions occurring or becoming sticky. Understanding these effects and
their implications for obsessions, if any, will require more complex models that
incorporate the interactions between various regions and neuromodulators.
Acknowledgments
Figures 13.A1 and 13.A2, from Friston et al. (2017), are used under the terms of the
Creative Commons Attribution 4.0 International License (CC-BY).
14
Psychological Interventions
as They Relate to
Intrusive Thinking
Intrusive, Emotional Mental Imagery
after Traumatic and Negative Events
Emily A. Holmes, Lisa Espinosa, Renée M. Visser,

Michael B. Bonsall, and Laura Singh
Abstract
Common across psychological disorders, intrusive, emotional mental images are
sensory-perceptual representations that intrude involuntarily into the mind. Mental
health treatments typically focus on entire disorders with multiple symptoms. This
chapter suggests focusing on core clinical symptoms (i.e., intrusive imagery). Existing
psychological therapy techniques (e.g., imagery rescripting) are promising, but under-
lying treatment mechanisms need to be better understood.
Precise treatments and preventions are required. Using the example of psychologi-
cal trauma, this chapter argues that psychological interventions can be developed in the
laboratory: effective experimental analogues of trauma can generate intrusions so that
putative interventions that modulate intrusions can be explored at various mechanis-
tic levels (e.g., molecular, cognitive, social). Examples of targeting “new” (i.e., Day 1
of the traumatic event) memories include a simple cognitive interference intervention
that holds promise for preventing intrusive images after trauma (a behavioral protocol
including Tetris game play). This intervention specifically targets intrusive involuntary
memories while leaving voluntary memory intact. Work on targeting “old” (as of Day 2)
memories is at an earlier stage. Research on reconsolidation update mechanisms appears
valuable in reducing older trauma memories via interference interventions, again with a
behavioral task interference technique. To understand mechanisms across different lev-
els (e.g., molecular, cognitive, or social), mathematical models can aid the identification
of causal mechanisms involved in memory formation. Questions are posed to instigate
discussion of future science-driven psychological interventions for intrusive images.
288 E. A. Holmes et al.
Introduction
What Is Intrusive Thinking?
One of the rare clinical psychology textbooks dedicated to intrusive thinking

(Clark 2005:4) states:
we define unwanted clinically relevant intrusive thoughts, images or impulses
as any distinct, identifiable cognitive event that is unwanted, unintended, and
recurrent. It interrupts the flow of thought, interferes in task performance, is as-
sociated with negative affect, and is difficult to control.
In this chapter, we focus on intrusive thoughts in the form of mental images.

This is because intrusive images occur across mental disorders. Further, it has
been shown that, compared to verbal thought, imagery has a more powerful
impact on emotion (Holmes and Mathews 2010) and thus carries the most
weight in psychopathology. We suggest that a focus on mental imagery opens
up novel angles for treatment innovation.
Mental images are sensory-perceptual representations; that is, like percep-
tion in the absence of percept (Pearson et al. 2015), as if “seeing in the mind’s
eye.” They can occur in any sensory modality, not just visual. Imagery can be
of past memories or simulations of future events. When images intrude invol-
untarily into the mind, they can carry strong emotion and influence behavior.
Intrusive image-based memories of a traumatic event are a core clinical
feature of both acute stress disorder and posttraumatic stress disorder (PTSD)
(American Psychiatric Association 2013). For instance, a person who expe-
rienced a traumatic road traffic accident may develop a distressing intrusive
image of a red truck coming toward them right before the accident happened
(Iyadurai et al. 2018). Mothers who have experienced traumatic childbirth may
see an intrusive image of the hospital lights above them when they went into
surgery (Horsch et al. 2017).
Intrusive images are not only present after trauma, such as in PTSD, they
can appear in numerous other psychological disorders (Holmes and Mathews
2010). For instance, people with social phobia may repeatedly see themselves
performing badly in a social situation, and such negative self-images play a
causal role in maintaining symptoms in social phobia. People with depres-
sion may experience intrusive images of themselves being rejected or so-
cially isolated, whereas they usually experience impoverished positive future
imagery (Hales et al. 2011; Newby and Moulds 2012; Holmes et al. 2016a).
Intrusive mental imagery may act as an emotional amplifier of all mood states
in people with bipolar disorder (Holmes et al. 2016b): Vivid negative mental
imagery, such as seeing oneself committing suicide in the future, may drive
despair. Vivid overly positive imagery, such as seeing others responding ex-
tremely well to one in social situations, may drive mania. These are only some
examples of intrusive images across mental disorders; the list could even be
Current Methods of Psychological Interventions 289
further extended to account for intrusive imagery in obsessive-compulsive dis-

order (OCD) (Coughtrey et al. 2015), body dysmorphic disorder (Osman et al.
2004), and agoraphobia (Day et al. 2004).
Here, we focus on intrusive mental imagery associated with psychological
trauma and depression. Thus, the mechanisms, treatments, and mathematical
framework discussed may or may not apply to other psychological disorders char-
acterized by intrusive images, such as OCD, which we do not address in detail.
What Are Psychological Interventions?
Psychological therapy is described as an interpersonal intervention, usually

provided by a mental health professional, such as a clinical psychologist who
employs any of a range of psychological techniques. Various schools of ther-
apy include cognitive behavioral therapy (CBT), psychoanalysis, systemic
therapy, and so forth. The CBT model was derived from a combination of
principles from behavioral and cognitive psychology alongside clinical experi-
ence. In brief, CBT focuses on challenging and changing unhelpful cognitive
distortions (e.g., thoughts, beliefs, and attitudes) and behaviors and improv-
ing emotional regulation. The individual learns personal coping strategies to
solve current problems. The therapist assists the individual in identifying strat-
egies to address goals and improve symptoms. CBT asserts that maladaptive
thoughts and behaviors influence the development and maintenance of psycho-
logical disorders, and thus symptoms can be reduced through new information-
processing skills and coping strategies.
Is there evidence that these methods actually work? Critically, how do we
choose which psychological or pharmacological intervention to use? Evidence-
based guidelines, such as put out by the U.K. National Institute for Health and
Care Excellence (NICE), critically review the full clinical literature and make
recommendations on the basis of how effectively treatments work. There have
now been hundreds of trials of psychological treatments. These can be classified
by mental disorder; in the case of intrusive thinking, PTSD provides an example.
From such a review of the clinical research evidence, the main recommen-
dation for PTSD, in terms of a first-line treatment, is individual trauma-focused
cognitive therapy (National Institute for Health Care Excellence 2018). This
is a tailored form of cognitive behavioral psychological therapy which fol-
lows a clear, validated manual; typically 8–12 sessions, each an hour long,
are held with a mental health professional who is highly trained in its specific
delivery. It includes elaboration and processing of the trauma memories, pro-
cessing trauma-related emotions, and restructuring trauma-related meanings
for the individual. The evidence-based guidelines for PTSD also recommend
another psychological treatment: eye movement desensitization and reprocess-
ing. CBT interventions can also be targeted at specific symptoms, such as sleep
disturbance or anger. Drug treatments are only recommended as a secondary
approach if patients prefer drugs over therapy (National Institute for Heath and
Care Excellence 2018), as the evidence is less strong for clinical effectiveness.
Drug treatments after trauma are not recommended as a preventive strategy; in
particular, benzodiazepines can worsen symptoms.
Since the inception of CBT in the 1960s, great advances have been made.
For instance, we now have highly effective evidence-based CBT treatments
for full-blown PTSD. This is one of the areas in which we have the best treat-
ments: some of the best PTSD trials have recovery rates of 75%, whereas the
norm in CBT is 50% (Holmes et al. 2014). However, there are still many ways
in which we need to improve psychological treatments (Holmes et al. 2018).
How can we adapt or simplify them to reach more people? How do we under-
stand the critical ingredients in a psychological treatment in so doing? Can we
also “prevent” rather than “cure”? That is, can we prevent intrusive memories
after trauma rather than only have treatments once the full-blown disorder has
been established?
When we think, we can think in the form of words (verbal thought) or men-
tal images (sensory representations in any modality such as visual, olfactory,
or auditory images). The dominant focus in psychological treatment research
and therapy, such as CBT, has been on verbal thoughts. A focus on mental im-
agery is one alternative and may open up opportunities for both research and
treatment innovation.
What Can We Do at the Moment?
CBT Techniques That Target Intrusive Mental Imagery
In a handbook for clinicians and patients, we have described four face-to-face

mental imagery-focused techniques in CBT: imagery rescripting, metacogni-
tive techniques, imagery-competing tasks, and enhancement of positive im-
agery (Holmes et al. 2019). In contrast to the CBT techniques that typically
focus on a whole disorder (rather than on one symptom), these techniques spe-
cifically focus only on distressing intrusive mental images. During imagery
rescripting, a person is asked to bring an intrusive negative image to mind and
describe it in detail (e.g., an image of oneself in a car crash). An alternative
image is then introduced to transform and update the negative image (e.g., an
image of oneself being well today despite the car crash). Metacognitive tech-
niques aim at lessening an image’s impact on a person by emphasizing that it is
not real (e.g., switching attention from the internal image to the outside world
or reinforcing the image’s unreality by making it look comical). Imagery-
competing techniques aim at disrupting or interfering with intrusive images
with a competing task (e.g., a visuospatial computer game) while the image is
active in a person’s mind. Positive imagery enhancement aims at encouraging
people to generate positive future images or repeatedly train them to interpret
ambiguous scenarios in a more positive way (Holmes et al. 2019).
Need to Understand Psychological Interventions in the Lab and

Be Able to Disseminate Them Globally
To understand how existing treatments for intrusive images work, we need to

clarify the key mechanisms driving these treatments. An experimental psycho-
pathology approach allows a direct test of whether an experimental manipula-
tion of a specific mechanism leads to a change in intrusions where key processes
that maintain or change aspects of psychopathology can be identified. Thus,
we need to take current treatments back to the laboratory and examine specific
treatment mechanisms modulating intrusions in isolation and under controlled
conditions. We can then remove irrelevant strategies from current treatments
and develop novel approaches that target the essential causal mechanisms
modulating intrusive images and are more effective and precise (Holmes et al.
2018). Such novel approaches need to be easily scalable to meet the global need
for mental health treatments. Thus, we need to develop simple, brief, and flex-
ible interventions that can be adapted to people’s needs across cultures. Such
interventions should ideally not require highly trained mental health profession-
als but should be able to be delivered by trained lay mental health care workers
via innovative and remotely accessible online platforms (Holmes et al. 2018).
What Do We Need to Do Next?
Research Paradigms to Study and Generate Intrusions
To improve treatments that target intrusive thinking, we need to be able to

generate intrusions and study their crucial underlying mechanisms in con-
trolled laboratory settings. Going back to the lab allows us to change focus
from complex real-life situations with clinical populations to simpler ex-
perimental procedures with nonclinical human (or nonhuman) populations
(Visser et al. 2018). That is, we need experimental models that incorporate
the dynamic nature of intrusive memories (Visser et al. 2018). Different para-
digms can be used to generate intrusions in the laboratory. Here, we focus on
two commonly used analogs of stressful events/trauma in anxiety and PTSD
research: fear conditioning (Pittig et al. 2018) and the trauma film paradigm
(James et al. 2016a).
Fear Conditioning
One well-known experimental method to investigate involuntary expressions

of aversive memory in both human and nonhuman animals is Pavlovian fear
conditioning. This paradigm has been used to investigate aversive associative
learning, considered to be an important mechanism in the etiology of anxi-
ety disorders (Pittig et al. 2018). Indeed, in real-life settings, learning what
is threatening or safe is important for survival. However, the association of
neutral cues with threat can become maladaptive when such conditioned fear
fails to extinguish long after the danger has passed or overgeneralizes to a
wider range of contexts.
Fear-conditioning paradigms allow for an investigation of the emergence,
persistence, and resurgence of maladaptive fear responses (Pittig et al. 2018). It
has been suggested that intrusion, hyperarousal, and hypervigilance symptoms,
which characterize PTSD, may arise as a result of conditioned fear responses
(for a review, see Norrholm and Jovanovic 2018). Note, however, that fear-
conditioning experiments have mainly investigated arousal or hypervigilance
(e.g., skin conductance, startle reflex–fear responses) that are mostly relevant
to anxiety disorders.
Intrusions, however, are the hallmark feature of PTSD, which is no longer
classified as an anxiety disorder but as a trauma- and stressor-related disor-
der in DSM-5 (American Psychiatric Association 2013). As fear-conditioning
paradigms do not specifically account for the image-based episodic nature of
intrusive thoughts (Visser et al. 2018), it remains an interesting open question
whether this paradigm could also be used to generate intrusive memories of
a stressor in the laboratory. A more ecologically valid and clinically relevant
experimental model may be needed to generate and study intrusions per se in
the laboratory.
Trauma Film Paradigm
In the trauma film paradigm, participants are asked to view a composition of

short, distressing film clips with traumatic content (James et al. 2016a). This
paradigm has been shown to induce intrusive memories to clips of the film
(i.e., with image-based episodic nature).
Trauma is defined as exposure to death, threatened death, actual or
threatened serious injury, or actual or threatened sexual violence (American
Psychiatric Association 2013). Notably, in addition to direct exposure (e.g.,
as a victim or witness), repeated or extreme indirect exposure to aversive
details of trauma, usually over the course of work (e.g., when a police of-
ficer has to view pictures of murder), is now included as part of the diagnos-
tic criterion for what comprises a traumatic event in the DSM-5 (American
Psychiatric Association 2013). This recent inclusion of indirect exposure
to trauma underscores the ecological validity of the trauma film paradigm
(James et al. 2016a).
Of note, intrusive memories can also be induced by overly positive film
stimuli (Davies et al. 2012) or depression-linked film material (Lang et al.
2009). Thus, the trauma film paradigm is not only useful for studying intrusive
images related to PTSD but also for studying intrusive thinking in depression
or bipolar disorder.
In studies using the trauma film paradigm, intrusive memories of the film
are usually monitored in a paper-and-pencil diary directly when they occur
over the course of daily life (James et al. 2016a). This method records intrusion
frequency data over longer time frames and carefully matches intrusions to the
trauma film (participants usually record whether they had an intrusion or not
for several time periods per day over the course of one week and briefly de-
scribe each intrusion’s content). Mixture models can be a useful tool to analyze
such diary data because they model intrusion and non-intrusion data differently
(see discussion on mechanisms and mathematics below for further details).
Additional Methods
Watching visual stills of distressing content (e.g., injured people) has also been
shown to generate intrusions two days later (Battaglini et al. 2016). In addition,
listening to negative arousing stories while watching a slide show of pictures
can generate negative emotional memories (Galarza Vallejo et al. 2019).
Levels of Mechanism to Modulate Intrusive Emotional Images
Using controlled and standardized experimental procedures and the possibil-

ity to focus on specific clinical targets is an essential step toward understand-
ing complex clinical disorders (Visser et al. 2018). Once one has successfully
generated intrusions in the laboratory, it is possible to study specific mecha-
nisms that could modulate them (e.g., reduce intrusion frequency or distress/
vividness of intrusions). At this point it is important to note that such intru-
sions can be modulated at any level of mechanism (e.g., molecular, cognitive,
or social). Here we discuss examples of paradigms modulating intrusions at
various levels: pharmacological approaches operating at the molecular level,
visuospatial interference interventions operating at the cognitive level, social
support operating at the social level, and other examples such as sleep and
wakeful rest.
Molecular Level: Pharmacological Approaches
Pharmacological approaches may offer a way to modulate intrusive memories.

For instance, inhibiting N-methyl D-aspartate receptor (NMDAR)-dependent
memory consolidation through antagonistic drugs may reduce the frequency of
intrusive memories after trauma. In line with this idea, inhaling the NMDAR
antagonist gas nitrous oxide (N2O) shortly after a laboratory analog of trauma
fastened the reduction of intrusive trauma-related memories compared to in-
haling medical air over the course of the following week. Of note, N2O led
to an increase in intrusion frequency in those individuals who were highly
dissociated at baseline, urging caution regarding the use of N2O in dissociated
individuals (Das et al. 2016).
Cognitive Level: Visuospatial Task Interference
Research in cognitive psychology and experimental psychopathology has

shown that cognitive interference interventions (memory orientation/reminder
cue and visuospatial task) may be a promising technique to reduce both the
frequency of intrusive images as well as the level of distress and vividness as-
sociated with them (Iyadurai et al. 2019). The working mechanisms behind this
intervention are based on three assumptions:
1. Intrusive memories can be altered shortly after an event or at retrieval
(Visser et al. 2018).
2. The capacity of people’s working memory is limited (Baddeley 2003).
3. Visuospatial tasks occupy working memory resources that would be
needed to (re)consolidate intrusive mental images (James et al. 2015).
Thus, engaging in a visuospatial task such as a visuospatial computer game
like Tetris (James et al. 2015; Iyadurai et al. 2018), or a complex finger
tapping exercise, at a time when mental images of the event are active,
may disrupt these distressing images. It is hypothesized that the intervention
works because the two processes compete for visual processing resources
and the brain cannot attend equally to the distressing image and the visuo-
spatial task. Importantly, such task interference has to take place at a time
when the memory is labile and vulnerable to alteration (McGaugh 2000;
Nader 2003).
Even though visuospatial interference interventions have mostly been in-
vestigated in relation to distressing trauma memories, they also work with
overly positive material (Davies et al. 2012). This suggests that the mecha-
nisms apply to intrusive emotional memories in a more general sense rather
than only to trauma-related intrusive images.
Social Level: Social Support
There has been an increased interest in the impact of social factors on emo-
tion regulation. Both human and nonhuman experiments have shown that the
presence of another during an aversive experience may work as a buffer by
reducing fear responses (Thorsteinsson and James 1999; Mikami et al. 2016).
Experiences of social support could increase the process of learning what is
safe in the environment (social safety learning) through social support interac-
tions, which in turn decrease stress reactivity to stressful experiences (Ditzen
and Heinrichs 2014).
After a psychologically traumatic event, social support (i.e., supportive
interactions with family and friends) is believed to be associated with hav-
ing fewer posttraumatic cognitions (e.g., trauma-related thoughts and beliefs),
which in turn is associated with PTSD symptoms (Woodward et al. 2015).
These results signal a need to investigate social interactions and social support
after a negative or stressful experience as a potential causal mechanism for the

development or maintenance of psychological disorders and to study this in
the laboratory. For example, if the absence of social support after trauma leads
to people having more intrusions, whereas perceived social support causes
people to have less intrusions, we could specifically target this mechanism in
future preventive treatments for people who experienced a traumatic event.
Studying the mechanisms at a social level may have relevance for public
mental health. Brief and low-intensity social support interventions may not
need to be delivered by highly trained professionals but could instead be im-
plemented by members of the public. Social prescription refers to the idea of
linking patients in primary care with sources of support within the community,
for instance, enabling health care professionals to refer patients to a service
provided by the voluntary and community sector alongside existing treatments
to improve health and well-being (Bickerdike et al. 2017). In line with the
idea of social prescription, social support interventions including emotional,
instrumental, and informational support could be delivered by volunteers (e.g.,
hospital volunteers) who are already present in many medical facilities, thus
allowing us to scale up preventive interventions.
Other Levels: Sleep or Wakeful Rest
An example of how memory could be boosted rather than blocked is wakeful

rest (Dewar et al. 2014). A brief wakeful rest period after learning may actu-
ally enhance memory in the short (after 15 minutes) and long term (after seven
days) compared to performing a nonverbal task (note that these studies tested
participants’ declarative memory). The wakeful-resting period is thought to
boost recently acquired memories by isolating the memory trace of the story
(or nonwords) from competing memories, making the memory easier to re-
trieve at a later stage (Dewar et al. 2014). These results confirm the crucial role
of the memory consolidation period in the strengthening of new memories,
here through spontaneous reactivation during wakeful resting.
What remains to be further explored is the possible involvement of similar
processes in the maintenance of intrusive thoughts during the consolidation
period of emotional material. Wakeful rest might actually be what trauma pa-
tients usually do when waiting for medical care in the emergency department
after a traumatic experience. Thus, investigating the effects of wakeful rest on
intrusive memories after a traumatic event could have clinical implications
and guide the development of future interventions. In line with this idea, a
few studies have already investigated the role of sleep and sleep deprivation
after trauma on intrusive memories (e.g., Porcheret et al. 2015, 2019, 2020).
As these initial investigations revealed mixed results, further research on the
role of sleep and wakeful rest as candidates to modulate intrusive memories is
clearly warranted.
Modulating the Frequency of Intrusive Memories: From Lab to Clinic
New Memories of a Traumatic Event
By “new” we refer to Day 1 of the (experimental or real) traumatic event.

Intrusive memories of trauma have a clear onset (i.e., the time of the traumatic
event), making them amenable to study. This allows us to investigate ways to
intervene with the initial memory consolidation of a problematic image before
it causes further distress at any of the above described levels of mechanism,
such as the molecular level (Das et al. 2016).
Several studies using the trauma film paradigm (James et al. 2016a) indicate
that after the experimental trauma (30 min or 4 hr), performing a brief cogni-
tive interference intervention (comprised of a memory orientation/reminder
cue, mental rotation instructions, and playing the visuospatial computer game
Tetris) reduces intrusive images compared to not performing any task (e.g.,
Lau-Zhu et al. 2019). Two proof of principle randomized control studies have
recently extended this effect to a clinical setting that involves (a) road traffic
accident survivors who are waiting in the emergency department (Iyadurai et
al. 2018) and (b) mothers who experienced traumatic childbirth (Horsch et al.
2017), both within the first six hours after the traumatic event.
Psychological interventions for traumatic memories should ideally inter-
fere with the involuntary, intrusive aspect of a memory but should not impair
voluntary memory expression (Lau-Zhu et al. 2019). A person who has experi-
enced sexual abuse by a piano teacher would, for instance, not want images of
the abuse to intrude on their mind involuntarily, whereas they may want to be
able to recall episodes and facts about the event when required for legal reports
(see Figure 14.1). Experimental studies in the laboratory make it possible to
investigate such a distinction. Findings suggest that a visuospatial interference
task intervenes with the involuntary (intrusive) memory, whereas the voluntary
memory remains intact when controlling for potential other task characteristics
(Lau-Zhu et al. 2019).
This data raises the intriguing possibility that intrusive image-based memo-
ries are in fact “special” and can be selectively targeted by visuospatial inter-
ference interventions, whereas voluntary memory remains unaltered (Lau-Zhu
et al. 2019). In contrast to traditional single trace theories of memory, which ar-
gue that involuntary and voluntary memories are derived from the same mem-
ory system, this data conforms to separate trace theories, stating that different
memory traces underlie involuntary and voluntary memories. Thus, intrusive
reexperiences may be supported by a specialized perceptual memory system
that is functionally dissociable from the episodic memory system support-
ing voluntary recall of the same event, in line with dual representation theory
(Brewin 2014). Visuospatial interference intervention (e.g., reminder cue and
Tetris) may then preferentially disrupt this sensory-perceptual memory sys-
tem, whereas the episodic memory system remains unaffected (Lau-Zhu et al.
Intrusive memories of Knowledge of facts or

fragments of the trauma trauma episodes
Unwanted emotion-laden Episodes and facts recalled
memories that spring to mind deliberately when choosing to
unbidden in form of sensory recount the trauma such as for
imagery such as sight, sounds, legal reports (e.g., address of
smells (e.g., seeing an image of abusive piano teacher; some
abusive piano teacher’s face and details of what happened
smell of aftershave) during the trauma)
Declarative
Voluntary
Conditioned responses to
trauma reminders
Attentional bias to threat; Preserve!
physiological reactions to internal
or external trauma clues (e.g.,
jumpiness when entering a
classroom; sweating when sitting
next to someone; muscle tension
when hearing a certain tune)
Nondeclarative
Involuntary Control!
Figure 14.1 Diagram depicting how different memory systems may represent various
aspects of a traumatic event (e.g., sexual abuse by a piano teacher). In general, clinically
beneficial interventions should aim to target the maladaptive involuntary expression of
trauma memories (e.g., intrusive memories) while preserving its voluntary recall (e.g.,
ability to testify in court). Adapted after Visser et al. (2018).
2019). In contrast to widely used fear-conditioning paradigms, the trauma film

paradigm may be particularly useful to assess these different aspects (readouts)
of trauma memory in the laboratory (Table 14.1). Future work is warranted.
Old Memories of a Traumatic Event
By “old” we refer to Day 2 onward of the (experimental or real) traumatic

event, and in one study many years later. Most work discussed above has fo-
cused on the time window that is thought to overlap mainly with (synaptic)
consolidation. Consolidation refers to a strengthening of local neural circuits
via a cascade of molecular processes involving protein synthesis and the forma-
tion of new synaptic connections necessary for a memory to persist in the long
term. As illustrated above, interventions delivered during this time period (i.e.,
minutes to hours after an event) are able to interfere with the newly formed
memory and reduce its intrusiveness (McGaugh 1966, 2000). Promisingly, re-
cent research suggests that, under the right circumstances, even established
memories can be modified. Rather than there being a one-off opportunity,
memories can, upon their retrieval, enter a transient labile state; that is, they
Table 14.1 Overview of different aspects of trauma memory that can be targeted and
associated research approaches, from animal models (bottom) to clinical populations
(top). Left: different levels at which trauma can be modeled. Middle: potential targets
for intervention. Right: memory readouts: (1) occurrence of intrusive images (e.g., diary,
provocation task), (2) event details (e.g., interview), (3) learning episode details (e.g.,
recognition test), (4) self-report of symptoms, (5) rating of subjective distress, (6) un-
conditioned stimulus expectancy, (7) attentional bias, (8) approach/avoidance behavior,
(9) noninvasive physiology, (10) invasive physiology. Note: voluntary memory recall
(e.g., trauma details) can be measured in humans but is not the key clinical target of a
treatment. Adapted after Visser et al. (2018).
Memory
Level Targets
readouts
Complex real-life emo- • Intrusive image-based memories or other 1, 2, 4, 9

tional memory, PTSD experiences (DSM5 cluster B)
patients: • Physiological reactivity to internal/external
• Heterogeneous, possible reminders (DSM5 cluster B)
multiple or sustained • Unwanted avoidance of internal/external
trauma exposure reminders (DSM5 cluster C)
• Clinical >1 month • Negative mood/cognition (DSM5 cluster D)
posttrauma • Hyperarousal (DSM5 cluster E)
• Functional impairment (DSM5 cluster G)
Simpler real-life emo- • Intrusive images 1, 2, 4,
tional memory, humans • Subjective distress 6–9
• Single trauma exposure • Unwanted avoidance
soon after event • Physiological reactivity
• Simple phobia
• Possibly subclinical Associated with (reminders of) specific real-
life situations
Complex experimental • Intrusive images 1, 3, 5, 9
emotional memory, • Subjective distress
humans • Physiological reactivity
• Trauma analog from
viewing of aversive film Associated with (reminders of) aversive lab
clips stimuli
Simpler experimental • Subjective distress 3, 5–9
emotional memory, • Avoidance
humans • Physiological reactivity
• Aversive conditioning,
still pictures paired to Associated with conditioned cues and
electric shocks contexts
Simpler experimental • Avoidance 8–10
emotional memory, • Physiological reactivity
nonhuman animals:
• Fear conditioning, tones Associated with conditioned cues and
paired to electric shocks contexts
can become malleable again (e.g., Sara 2000; Alberini 2005; Nader and Hardt
2009). The restabilization of a memory is a putative process termed memory
reconsolidation (Nader et al. 2000). This process is dependent on de novo pro-
tein synthesis; interventions that directly or indirectly target this process thus
have the potential to change maladaptive emotional memories (Milton and
Everitt 2012), including those giving rise to intrusive images of trauma. Figure
14.2 depicts different time windows of memory malleability.
Different interventions can interfere with the reconsolidation of a memory
on different levels. On a molecular level, fear-conditioning studies in rodents
have shown the potential of pharmacologically disrupting one-day-old (Nader
et al. 2000; Lee et al. 2006; Ortiz et al. 2015) and even one-month-old (Gräff
et al. 2014) memories, resulting in a persistent attenuation of conditioned fear
responses. Even though less consistent (Lonergan et al. 2013), these types of
findings have been translated to studies in humans, where the beta-adrenergic
antagonist propranolol was used to disrupt pharmacologically one-day-old fear
memories (Kindt et al. 2009) as well as much older memories; that is, fear
memories that underlie simple phobias for spiders (Soeter and Kindt 2015). On
a cognitive level, behavioral interventions, including extinction after memory
retrieval procedure, have been shown to attenuate one-day-old fear memory in
rodents (Monfils et al. 2009). This finding has been translated to one-day-old
(Schiller et al. 2010) and one-week-old (Steinfurth et al. 2014) memories in
humans, and more recently also to older memories such as those underlying
simple phobia for spiders (Björkstrand et al. 2016) and snakes (Telch et al.
2017). For further details, we refer the reader to recent overviews on mem-
ory reconsolidation literature (Lee et al. 2017; Elsey et al. 2018; Monfils and
Holmes 2018).
With regard to intrusive memories, a visuospatial interference intervention
administered after a reminder cue was effective in reducing intrusive memories
for established (24-hour-old) memory of experimental trauma (James et al.
2015). In this study, individuals who underwent a memory reactivation pro-
cedure and performed an intervention, including Tetris game play, had fewer
intrusive memories than a no-reactivation/no-Tetris group. More recently, two
studies used a similar reactivation and cognitive task interference procedure,
administered three days (Kessler et al. 2020) or four days (Hagenaars et al.
2017) after trauma film viewing; again, a reduction in subsequent intrusive
memories was demonstrated. While both studies showed that an active control
condition (verbal task) also reduced intrusions compared to a no-task control,
in one study the effect was significantly larger for the visuospatial interfer-
ence intervention compared to a verbal control task (Kessler et al. 2020).
Interestingly, and again in line with separate trace theories (Lau-Zhu et al.
2019), both Kessler et al. (2020) and James et al. (2015) showed that the inter-
vention left voluntary memory (i.e., performance on a recognition task) intact.
Still, more work is warranted.
300
Encoding Minute / hours

Pre 1 day post Weeks / months / years post
episode post
Consolidation Systems consolidation Windows of oportunity

for memory updating
Retrieval Retrieval
Memory susceptibility
Time
Unsuccessful Successful Successful intervention (blockade soon Unsuccessful Successful
intervention intervention after encoding, within the encoding intervention (blockade intervention
(pre-trauma) (blockade during context, or preceded by cue that only, no retrieval cue, (blockade preceded
encoding) directs internal attention to trauma) or vice versa) by retrieval cue)
Figure 14.2 In the hours after an experience, memories are believed to go through an initial labile phase before being stored into stable long-term
memory (i.e., consolidation). The purple arrow depicts different time intervals with respect to the encoding of an aversive episode. The gradients
below indicate the putative processes of memory encoding and consolidation that occur during these different intervals. Recent insights suggest
that certain aspects of memories, including the intrusiveness, are not necessarily permanent. Instead, they may become transiently malleable upon
reactivation, rendering them susceptible to interference or updating before returning to a fixed state, a process referred to as “memory reconsolida-
tion.” This offers a second window of opportunity to interfere with consolidated memories (shown as yellow background shades), making them
less intrusive. Successful interventions (blue arrows) need to be timed such that the blockade interferes with memory when it is in an active, sus-
ceptible state (indicated by the dotted yellow line), either in the first hours after trauma or at later time intervals after a retrieval procedure (e.g.,
reactivation through reminder cues). In the first hours after an experience, blockade procedures may also need to be preceded by cues that orient
attentional resources to the event in order for procedures to successfully interfere with it (e.g., when the intervention is delivered in a context other
than the one in which the trauma occurred). Unsuccessful interventions, timed when memories do not yet exist or are in a fixed state (i.e., not
recently retrieved), are shown. Adapted after Visser et al. (2018).
Real trauma memories are typically stronger and broader than aversive
memories formed in the laboratory. Finding the optimal conditions and re-
minder cues to reactivate and render a memory labile (a first step for suc-
cessful interference) is assumed to be much more challenging (Monfils and
Holmes 2018) for real memories of trauma. Yet, a recent study on inpatients
with complex trauma (Kessler et al. 2018) has shown promise in attenuat-
ing the intrusiveness of memories for old trauma some years previously.
Twenty patients monitored the occurrence of intrusive trauma memories over
the course of their admission (5–10 weeks). Weekly interventions involved a
memory reminder for a selected (particularly distressing) memory, followed
by 25 minutes of playing Tetris. A within-subjects multiple baseline design
was used, in which the pre-intervention period was varied. Further, some in-
trusions were never targeted by the intervention. The frequency of targeted
intrusive memories reduced, on average, by 64% from baseline to the pos-
tintervention phase, whereas never-targeted intrusions reduced in frequency,
on average, by 11% over a comparable time period. This shows that even
persistent, older memories of real-life trauma can be changed using memory
interference techniques.
Despite its clear promise for clinical translation, it should be noted that a
number of potential limitations and boundary conditions of reconsolidation-
based clinical applications have been raised (Treanor et al. 2017; Monfils and
Holmes 2018). Moreover, at present, it is not possible to attribute conclu-
sively therapeutic gains to reconsolidation mechanisms (Elsey et al. 2018).
Nevertheless, the notion of memory plasticity has proved useful in inspiring
new avenues for intervention for older memories of trauma (Figure 14.2). Of
particular interest to our current discussion is the potential to modify intrusive
features of memory during time windows of memory plasticity.
To be able to interfere with older trauma memories, the memory trace has to
be activated in working memory via a reminder cue. According to reconsolida-
tion theory, there is an optimal duration for a reminder cue. When memory is
retrieved via a brief learning experience (e.g., one unreinforced conditioned
stimulus) it enters a labile state. However, if retrieval is prolonged (e.g., four
unreinforced conditioned stimuli), the memory might enter a “limbo state,”
and if it is prolonged further, finally extinction. In short, if the reminder cue
“dosage” (duration, instances) is too little, nothing happens (no labilization);
if it is too big, the memory may enter a “limbo state” (nothing happens) or
extinction learning—a new inhibitory trace is formed, fear/distress diminishes,
but this effect may be temporary as it does not alter the original emotional
memory trace (Lee et al. 2006; Merlo et al. 2014; Sevenster et al. 2014). All
three possibilities (no labilization, limbo state, extinction learning) are differ-
ent than the reconsolidation state, so the optimization of the retrieval procedure
follows an inverted U-shape.
From a clinical perspective, experiencing an intrusion may even offer an
opportunity to interfere with reconsolidation of this memory by engaging in
a competing cognitive task (e.g., playing Tetris) within a specific time frame
after the intrusion occurred (minutes). However, the question is whether spon-
taneous retrieval by means of experiencing an intrusion induces the required
reconsolidation state, or instead any of the other states, in which case one can-
not really interfere with it. This is an important empirical question which has
yet to be tackled.
Recently, an experimental paradigm has been developed to capture intru-
sive memories as they occur in the fMRI scanner (Clark et al. 2015). After
viewing scenes of traumatic events (trauma film paradigm; James et al.
2016a), particular scenes then intrude for an individual. A specific intrusive
memory is triggered in the scanner by a reminder cue. The first results of
experiencing an intrusive memory are shown in Figure 14.3. Understanding
the neural mechanisms of experiencing an intrusive memory may yield
insights for treatment (e.g., for neuromodulation strategies that could be
combined with behavioral interference techniques, such as our Tetris proce-
dure). Colleen Hanlon (this volume) discusses transcranial magnetic stimu-
lation (TMS). It is possible that understanding the neural mechanisms of
an intrusive event (e.g., Clark et al. 2015) alongside associated multivoxel
pattern analysis (e.g., Clark et al. 2014b) will inform how best to apply
TMS during an interference procedure (e.g., Kessler et al. 2018) to reduce
the occurrence of intrusive memories. However, to date this has not been
attempted.
Mechanisms and Mathematics
Mechanisms of cognition operate across the scales of brain organization

(Bonsall et al. 2015). If multiple processes operate at different scales of or-
ganization in psychopathology (e.g., posttraumatic stress reactions or mood
instability), aggregating the collective molecular and neuron interactions to
higher levels of organization (such as the network level or cognitive level)
might provide novel, emergent insights into the patterns associated with brain
function within and among individuals.
Using mathematical approaches to scale (appropriately) across a hierar-
chy from cognitive and emotional processes through neuron firing patterns
to candidate, molecular processes allows development of a mechanistic ap-
proach to cognition. This mechacognitive approach (i.e., using mathematical
approaches to move down a hierarchy from symptoms to candidate, molecu-
lar processes) may allow insights through a mechanistic approach to cogni-
tion (Holmes et al. 2016b). By developing descriptions of psychopathology, it
can also lead to novel approaches to understand the underlying neuroscience
of brain function.
Memory Formation, Consolidation, and Reconsolidation

Is a Probabilistic Process
As discussed above, memory consolidation and memory trace are not fixed
(Müller and Pilzecker 1900; McGaugh 2000). Neural changes and reorganiza-
tion operate through the time period from the perception of an event to the later
memory retrieval. Given that memory trace is probabilistic (changes between
states) and dynamic in time (and across spatial organization in the brain), we
develop mechanistic frameworks to link across scales of organization to cogni-
tion. While the molecular basis of memory involves epinephrine and cortisol
and protein synthesis to affect changes in synaptic consolidation (Dudai 2004),
scaling this up to focus on the neural mechanism of systems-level consolida-
tion requires appropriate tools. We argue that this is best achieved within the
frameworks and architectures of mathematics.
To bridge the gap between (intrusive) memory consolidation and treatment
through a mechacognitive lens, we use a mathematical framework coupled to
data analysis. Here, our framework focuses on intrusive memory consolidation
and cognitive interference interventions including a visuospatial task (Figure
14.4). In this way we consider how perception (bringing to mind) of a trau-
matic event (zM) following an orientation cue (also called reminder cue for
consistency with the reconsolidation literature) might lead to memory con-
solidating into an intrusive memory (iM) or a more neutral task memory (tM)
following a task (T).
To formalize this, we consider a discrete-time Markov chain (Kemeny and
Snell 1960) where memory moves through a series of states, and the probability
of moving from one state to the next is only dependent on present state. Discrete-
time Markov chain models underwrite many common state space models of
data (e.g., latent variable models, hidden Markov models, and Markov decision
processes). In our current (illustrative) application, we assume that the states
are directly observable and focus on the probability transitions from one state
to another and their implications for understanding therapeutic interventions.
Markov chains can be described by a directed graph where edges are the proba-
bilities of moving between states and vertices represent the states (Figure 14.4).
The directed graph illustrates the steps needed to move from trauma memory
state to consolidated task memory state. For instance, this approach allows clear
assumptions to be formulated; one assumption is that trauma memory needs to
be in a labile state before tasks can be undertaken and affect consolidation of the
task memory. We emphasize that this is all a probabilistic process as we learn
how to scale up to aggregate memory processes, driven from the molecular,
short- or long-term scale to the system level (Albo and Gräff 2018) to a level of
organization at the cognitive scale.
As noted, in this exemplar we consider four states: initial trauma memory
state, labile memory, intrusive memory, and consolidated neutral task memory.
This can be represented by the following transition matrix (N):
304
(a) 0 to 3 seconds (Intrusive memory involuntary recall > Control button press)
1.7 3
3 to 6 seconds (Intrusive memory involuntary recall > Control button press) 1.7 3
6 to 9 seconds (Control button press > Intrusive memory involuntary recall) 1.7 3
Right Left Right Left Right Left Right Left Right Left Right Left
Z = –16 Z = –4 Z=8 Z = 20 Z = 32 Z = 44
(b) Intrusion Control
Middle frontal Superior frontal Operculum
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 3 6 9 3 6 9 –3 3 6 9
–0.2 –3 –0.2 –3 –0.2 0
0
% BOLD signal change

–0.4 –0.4 –0.4
–0.6 –0.6 –0.6
Time (s) in relation to button press (0) Time (s) in relation to button press (0) Time (s) in relation to button press (0)
Left IFG Precentral Control > Involuntary recall

0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0
0 3 0 0
0 3 3
–3 6 –3 6 9 –3 0 6 9
–0.2 9 –0.2 –0.2
% BOLD signal change

–0.4 –0.4 –0.4
–0.6 –0.6 –0.6
Time (s) in relation to button press (0) Time (s) in relation to button press (0) Time (s) in relation to button press (0)
Figure 14.3 Intrusive memory involuntary recall. Top: Whole-brain analysis showing the increased blood oxygen level-dependent (BOLD)
response for intrusive memory involuntary recall versus control button press group at the two time bins (0–3 s and 3–6 s in relation to the button
press): note the significant differences in activation and the one time bin (6–9 s) of increased BOLD response for the control button press group
versus intrusive memory involuntary recall. Bottom: Region-of-interest profile plots of the signal change observed across each time bin from –3
to +12 s in relation to the button press. Intrusive memory involuntary recall signal change activation is shown in red; control button press signal
change activation in blue. Values are means; standard deviations represented by vertical bars. IFG: inferior frontal gyrus. Adapted after Clark et
305
al. (2015).
0 p1 1 p1 0
p3 p2 1 p2
0
N 1 p3 (1 p3 ) 1 p3 , (14.1)
0 0 1 0
0 0 0 1
where pij is a transition probability (moving from a state in row i to a state

in column j) such that rows of the matrix sum to one (by normalizing across
the transition probabilities within a row). p1 is the transition probability from
trauma memory state to labile memory state and, in our context, is the condi-
tional probability that memory is a labile state (lM) given a reminder cue (rC),
(Pr(lM | rC)). The probability that the reminder cue fails and intrusive memory
forms for initial trauma state is 1 – Pr(lM | rC). Here, we assume that this is a
logistic function, 1/(1 + exp(–α)), where α represents the strength of the re-
minder cue). Similarly, p2 is the transition probability from the labile memory
state to consolidated iM. In our context this is represented as a conditional
probability that a task intervention affects the formation of intrusive memories
(Pr(iM | T)), and 1 – Pr(iM | T) is the probability that task intervention is effec-
tive and leads to a consolidated neutral task memory. Here we assume that
p3
p1
Trauma state Labile memory
1 – p1 p2 1 – p2
Consolidated Neutral task

intrusive memory memory
p = 1.0 p = 1.0
Figure 14.4 Markov chain model: Directed graph representing a Markov chain
framework for exploring intrusive and more neutral task memory consolidation. Ar-
rows (edges) represent transition probabilities between states, and boxes (nodes) rep-
resent different memory states. p1 represents the probability that the reminder cue is
successful. We describe this probability as 1/(1 + exp(–α)), where α is the strength of
the reminder cue. p2 represents the probability that task intervention is unsuccessful and
parameterized here as 1/(1 + T), where T is the strength of the task intervention. p3 is the
probability maintained in a labile state. Once memories enter an intrusive memory or
neutral task memory they are fixed in these states (p = 1.0).
Pr(iM | T) = 1/(1 + T), where T is a measure of task strength. p3 is the probability

that a memory is maintained in a labile state. With these assumptions and the
stochastic process framework, to investigate the role of task interventions once
a memory is either in an intrusive or task consolidated state, the memory is
fixed (absorbed) into this state and not amenable to further alteration. While
memory consolidation and reconsolidation are dynamic processes, here we fo-
cus on the role of task interventions in affecting memory states.
Analysis of the Markov chain allows a number of different metrics to be as-
sessed including (a) expected time in a state, (b) variance of time in a state, (c)
dynamics over finite time steps, and (d) probability of absorption into a final
state as a function of covariates (e.g., Kemeny and Snell 1960).
Sensitivity analysis highlights the importance of particular probability tran-
sitions (or more specifically drivers of probability transitions) and the factors
influencing memory consolidation at different levels of organization. Here,
analysis of the Markov chain reveals that the labile memory state is transient.
Expected time in this labile memory state is a function of reminder cue prob-
ability (p1) and probability of staying in the labile memory state (p3) (Figure
14.5a). Increases in the probability of keeping the memory labile (p3), and in-
creases in reminder cue strength (α) lead to longer expected times of memory
in a labile state. However, the uncertainty (variance) in the length of time a
memory is labile is most likely influenced by the probability of keeping the
memory labile (p3) rather than reminder cue strength, and most uncertainty in
the length of time a memory is labile is greatest for low reminder cue strengths
and high probabilities of keeping the memory labile (Figure 14.5b).
Potentially more important is the probability that intrusive or task memo-
ries consolidate; that is, since this is an absorbing Markov chain with two end
states, whether memories persist in the iM or tM state. Analysis of the Markov
chain reveals that the probability that an iM consolidates following trauma is:
Pr iM | trauma 1 p p1 p2 . (14.2)
Equation 14.2 expresses the probability that an iM consolidates given that the
reminder cue fails (1 – p1) or the probability that the reminder cue is successful
and that the task intervention is not successful (p1 p2). For our parameterization,
where p1 = 1/(1 + exp(–α)) and p2 = 1/(1+T ):
1 1
Pr iM | trauma 1
1 exp 1 exp 1 T
(14.3)
1 exp T
.
1 exp T exp T
The terms that contribute to this conditional probability can also be determined
intuitively by looking at Figure 14.4 and tracing the paths from the trauma
state to the intrusive memory state, accumulating the probabilities along all
(a) (b)
1.0 1.0
labile memory states (p3)
Probability of staying in
labile memory states (p3)

0.8 0.8
Probability of staying in
2.0
1.75
0.6 0.6
1.50 1.5
1.25
0.4 0.4 1.0
1.00
0.5
0.75
0.2 0.2
0.50
0.0 0.0
0 1 2 3 4 5 0 1 2 3 4 5
Strength of reminder cue (Į) 6WUHQJWKRIUHPLQGHUFXHĮ
Figure 14.5 Markov chain analysis of transient memory states. (a) Expected times
and (b) variances in expected times in labile memory states in terms of strength of
reminder cue (α) and the probability of staying in the labile memory state (p3). Strong
reminder cues and high probability of maintaining a memory in a labile state favor long
retention times in this transient (labile) memory state. However, most uncertainty in ex-
pected retention times is observed for low reminder cue strengths and high probability
of maintaining a memory in a labile state.
direct and indirect paths. The marginal change in the probability of iMs is more
sensitive to changes in the probability of task success (p2) than the reminder
cue success probability (p1), as the following inequalities for the change in
d(Pr(iM | trauma)/dpi can be shown to hold:
d Pr iM | trauma d Pr iM | trauma
p1 p2 1 . (14.4)
dp2 dp1
The above inequality holds from our parameterization of p1 and p2:
1 1
1 , (14.5)
1 exp 1 T
which, again, are the probabilities along the path from trauma to task memory
consolidation (Figure 14.4). The probability that a task memory consolidates is:
Pr tM | trauma p1 1 p2 , (14.6)
which is the probability that the reminder cue is successful (p2), and the task
intervention is successful (1 – p2). Again, the terms that contribute to this con-
ditional probability can also be determined intuitively from Figure 14.4, trac-
ing paths from the trauma state to the consolidation state, and accumulating the
probabilities along all direct and indirect paths. Two key predictions from this
analysis are as follows:
1. The probability of maintaining a memory in a labile state (p3) has no

effect on the probability of iMs or tMs consolidating.
2. Most importantly, the probability of a successful reminder cue is criti-
cal: if this cue is unsuccessful (p1 tends to zero), then the probability
of an intrusive memory consolidating tends to 1 and the probability of
consolidated task memory tends to 0.
Formulating memory consolidation as a stochastic process, aggregating across
scales of organization, allows systems-level mechanisms associated with cog-
nition to be investigated. Next, we briefly show how the model can be vali-
dated against empirical observations and/or experiments.
Model Validation: Statistical Approaches
Determining the accuracy and applicability of a mathematical framework cen-

ters around model validation. Validating a model involves appropriate param-
eterization, goodness of fit to data, uncertainty quantification, and prediction.
Linking a mechanistic model, such as our Markov chain, to data involves sta-
tistics and statistical modeling.
While full model validation is beyond the scope of what we present here,
we show ways in which Markov chains can be parameterized from trauma
and iM studies and the likely predictions that arise from this parameterization,
deriving the conditional probabilities that (a) a reminder cue places a memory
in a labile state and (b) task intervention affects the probability of intrusive
memories has been approached empirically (e.g., James et al. 2015; Iyadurai et
al. 2018, 2019; Lau-Zhu et al. 2019; Visser et al. 2018).
To determine the efficacy of a reminder cue, a binary regression is needed
with probability of a successful reminder cue as a response with a set of ex-
planatory covariates. To determine the conditional probability that task affects
probability of memory consolidation, we have advocated appropriately ad-
dressing statistical issues, such as correlation structures (James et al. 2015;
Iyadurai et al. 2018) and/or heterogeneity (Iyadurai et al. 2019). One way to
derive appropriate probability estimates on the efficacy of a task is through the
use of mixture models (Cameron and Trivedi 2013), where statistical model-
ing of zero and nonzero intrusive memories (from diary data) is considered
differently.
Mixture models represent the nonzero and the zero observations separately
with two statistical models. First is to ask: Are the zero counts generated be-
cause of the iM/trauma process or something else (perhaps to do with data
collection)? This is a Bernoulli process with a probability (say p, where this
probability is to be determined by the set of explanatory variables) that the
zeros are generated by alternative processes than those under observation. So
(1 – p) is the probability that the zeros are generated by the iM/trauma process.
As iMs are count data (e.g., the number of intrusive memories per day), the
nonzeros are modeled assuming Poisson errors. The probability of j intrusive

memories can then be represented by the following mixture model:
p 1 p exp u if j 0
Pr iM j iM
u exp u . (14.7)
1 p if j 0
iM !
The mixture model applies a binary/binomial regression to determine p (top

line, right-hand side of Equation 14.7) and a Poisson regression to model the
nonzero counts (bottom line, right-hand side of Equation 14.7).
The binomial and Poisson components of the regression can consider dif-
ferent covariates (e.g., task/no-task, task strength, task quality) to determine
probability of intrusive memories and appropriately address heterogeneity
generated by an overinflated number of zeros.
To illustrate this approach for data analysis and application to the Markov
chain model, a zero-inflated Poisson model analysis was undertaken on a group
of patients involved in traumatic road traffic; these patients took part in a cog-
nitive interference intervention that included a reminder cue and a visuospatial
task (for details, see Iyadurai et al. 2018). This analysis reveals that the overall
probability (across this group of patients) of a successful intervention (1 – p2)
and no iMs is 0.542 (Figure 14.6). Together with the Markov chain analysis,
this empirical estimate of no iMs (i.e., zeros), given a cognitive task, predicts
that the probability of a consolidated task memory across this group of patients
would range from 0 to 0.542 (depending on the probability of a successful
reminder cue).
The opportunities for using mathematical approaches for linking across
mechanisms of cognition, different illnesses and traumata, modalities of per-
ception, and individual patients is a nascent approach. However, we believe
this mechacognitive approach has value along the continuum from the basic
through to clinical aspects of neuroscience and will provide a fuller under-
standing of memory consolidation.
Conclusions
Recent findings on intrusive mental images, reviewed in the first part of this
chapter, as well as the mathematical model on intrusive memory consolidation
and visuospatial interference interventions, introduced in the second part, give
reason to take a step back and think about how current psychological interven-
tions might be improved. We have proposed that to progress in this regard, we
need to know more about specific processes involved in intrusive thinking and
adopt a targeted treatment approach (Iyadurai et al. 2019). We conclude by
raising the following questions to invigorate discussion on how we can make
future psychological interventions more precise and effective.
(a) Placebo control group (b) Cognitive intervention group
0.5
0.20
0.4
0.15
0.3
Density
Density
0.10
0.2
0.05 0.1
0.00 0.0
0 5 10 15 20 0 2 4 6 8 10
Number of intrusive memories Number of intrusive memories
Figure 14.6 Density distribution of intrusive memories (adapted after Iyadurai et al.
2018) for (a) the attention placebo control group and (b) the active cognitive interven-
tion (reminder cue + visuospatial interference task) group. Statistical analysis on (b),
using a zero-inflated Poisson intercept-only model (Equation 14.7), reveals that the
nonzero counts have an intercept that is significantly different from zero (intercept =
0.908 ± 0.144 (SE), z-value = 13.72, p < 0.001), while the zero counts do not (intercept
= 0.0006 ± 0.144, z-value = 0.0004, p = 0.997). The expected probability of no intrusive
memories in the active intervention group, determined from Equation 14.7 and across
this group of patients, is 0.542.
What if we specifically target intrusive thinking as a primary outcome

rather than a multitude of symptoms of a given disorder? Intrusive im-
agery is common across psychological disorders and appears to be an im-
portant transdiagnostic factor regardless of specific diagnosis (Iyadurai et
al. 2019). By looking specifically at intrusive imagery rather than broad
and fuzzy assemblies of symptoms clusters, we may be able to radically
change current treatment approaches in mental health. For instance, while
we may not be able to treat PTSD or depression reliably as a whole (espe-
cially at scale), we may be able to target a specific issue, such as intrusive
memories, that is common to both disorders. This could also cross-fertilize
treatment approaches across different disorders. Interventions targeting in-
trusive memories of trauma might inform us about potential methods to
target intrusive thoughts in depression, as has been the case with imagery
rescripting. Developing a simple intervention that is precise and can be de-

livered exactly to the intrusion trace may be a galvanizing aim for experi-
mental medicine across disorders.
What if we can target involuntary intrusive memories yet leave voluntary

memory intact? One of the most heated debates in the literature on intru-
sive memories involves whether involuntary and voluntary memories of an
event are best represented by single or separate memory trace accounts (Lau-
Zhu et al. 2019). Based on a recent series of carefully controlled experiments
investigating this question, we argue that memory is in fact dissociable, and
involuntary intrusive memories stand out from voluntary memories. This
dissociation has important implications in clinical settings as well as for
society. We need to develop psychological interventions that can prevent
involuntary distressing images from intruding on one’s mind while still en-
abling people to voluntarily recall information about the event (e.g., to be
able to testify regarding a traumatic event in a court of law). Visuospatial
interference interventions are promising because they appear to target selec-
tively and precisely the intrusive memory trace while leaving the voluntary
memory intact (Lau-Zhu et al. 2019). A successful recovery posttrauma from
a clinical perspective is being able to talk about the traumatic event(s) when
one decides to or needs to, but not to have them continually intrude in one’s
mind against one’s will.
What if we are able to prevent new intrusive memories as well as tackle older
ones? In addition to preventing the consolidation of “new” intrusive memo-
ries with visuospatial interference interventions directly after a traumatic event
(the same day), we argue that a similar type of intervention could be adapted
to target “old” intrusive memories. Importantly, when targeting older trauma
memories (24 hr to several years after the traumatic event), studies have indi-
cated that an approximately 10-minute gap has to be added between reminder
cue and intervention, supposedly to make the memory trace malleable (Agren
et al. 2012; Schiller et al. 2013; James et al. 2015; Kessler et al. 2018, 2020),
although more research on this gap is needed. Methodologically, this opens ex-
perimental designs to study visuospatial interference interventions. For exam-
ple, where there are older intrusive memories of several different events, one
could target single intrusive images one after the other (i.e., one at a time) and
compare frequency and distress of targeted and nontargeted intrusions over
time (Kessler et al. 2018). Targeting old intrusive memories is clearly impor-
tant for vulnerable patient groups (e.g., refugees). Creating rational approaches
that look more like computer game play may be useful for those who do not
seek traditional psychological help because of perceived stigma. Developing
a brief, easily accessible, and nonstigmatized cognitive intervention that could
potentially be self-administered would fill an important gap to reach such vul-
nerable patient groups (Holmes et al. 2018).
What if we could go global and target large sections of populations suffering

trauma? Trauma is a global health issue, and to address this we need in-
novative interventions that can be scaled up to overcome current barriers in
psychological treatment. For instance, it is impossible for traditional psycho-
logical treatments to be delivered to the large number of people who need them
globally due to the lack of trained psychologists. Thus, we need interventions
that are of low intensity, require few resources to deliver, are culturally adap-
tive, and accessible to many (Iyadurai et al. 2019). Approaches such as the
cognitive interference intervention (if shown effective in large-scale clinical
trials) could potentially be readily delivered by nonspecialists or even be self-
administered. Thus, an important aspect of bringing an intervention to a global
level is how to train people successfully with different background knowledge
in delivering the intervention (Holmes et al. 2018). Relatedly, at the global
level, prevention may ultimately be as important as a cure. Rather than treating
psychological disorders after they have developed and caused burden on the
individual and society (in terms of suffering, health care costs, and loss of work
force), selective prevention for high-risk groups (e.g., firefighters, paramedics,
emergency department staff, war survivors) or universal prevention (i.e., ev-
eryone experiencing trauma could be treated no matter if they would actually
develop intrusive images or not) would be a useful way to combat maladaptive
intrusive thinking.
What if we could prevent intrusive memories as well as boost positivity at the

same time? We need to find a way to make interventions as effective as pos-
sible while keeping them simple. Rather than simply aiming for a reduction of
intrusive memories, we might also want to boost positivity, if we could find
a simple way to combine this within the same task procedure (e.g., increase
optimistic mental images of the future and direct attention to adaptive infor-
mation; see Kress and Aue 2017). Positive imagery and optimism can be two
main targets, potentially at once. Notably, the mathematical model on memory
(re)consolidation and task interference introduced here could help find the best
way to test this (and related) questions. For instance, the model raises the in-
triguing possibility that, at least in some situations, two weak tasks can lead
to a stronger outcome than one strong task, and this may help clarify at which
point in time specific interventions are most effective.
Acknowledgments
EAH receives grant support from the Oak Foundation, The Lupina Foundation, and the
Swedish Research Council (2017-00957). RMV is supported by a Netherlands Organi-
zation for Scientific Research (NOW Veni) grant (016.195.246). Figures 14.1 and 14.2,
from Visser et al. (2018), are used under the terms of the Creative Commons Attribution
4.0 International License (CC-BY).
15
Pharmacological
Interventions as They Relate
to Intrusive Thinking
Harriet de Wit and Anya K. Bershad
Abstract
Intrusive thoughts are features of numerous psychiatric disorders. They vary widely in
form, duration, frequency, and severity. They are associated with disorders with widely
differing pathophysiology, and they are likely to respond to different pharmacological
treatments. It is possible that intrusive thoughts represent a cross-diagnostic symptom
that can be a pharmacological target in their own right, separate from the associated dis-
order. This chapter considers the challenges in studying intrusive thoughts as a separate
entity. It examines intrusive thoughts that are symptoms of several different psychiat-
ric disorders and reviews the medications that have been used to treat them. It holds
that relatively little is known about the effects of psychiatric medications on intrusive
thoughts, either within disorders (separate from other symptoms) or across disorders. A
wide range of medications is used to treat intrusive thoughts that target different neu-
rotransmitter systems. In addition to the psychopharmacological armamentarium, new,
single dose treatments (e.g., ketamine, psilocybin, and MDMA) have emerged that may
specifically address intrusive thoughts across the psychiatric spectrum. In conclusion,
possible directions are discussed for identifying subcategories of intrusive thoughts that
could advance research and treatment in this area.
Introduction
While most of the time we have a sense of control over our own thoughts,
sometimes it seems as if our thoughts control us. Under these circumstances,
thoughts appear to assert themselves into our consciousness, influencing mood
and sometimes behavior. These unwanted images and urges, or “intrusive
thoughts,” can interfere with everyday functioning (Clark 2005). In an early
definition, Rachman (1981) identified three criteria for intrusive thoughts: (a)
the thought interrupts ongoing activity, (b) the thought is recognized to be
316 H. de Wit and A. K. Bershad
of internal origin, and (c) the thought is difficult to control. When intrusive
thoughts become severe enough to interfere with normal function, they be-
come clinically significant psychiatric symptoms and thus potential targets for
treatment. In this chapter we review several psychiatric disorders in which in-
trusive thoughts are a prominent symptom and summarize the medications that
have been used to treat these disorders. We also comment on future directions
of pharmacological treatments for this symptom.
In the context of psychiatric disorders, intrusive thoughts vary widely in
their nature, form, frequency, and controllability, and this variance can make
them difficult to study as well as to treat. Intrusive thoughts may occur as part
of normal human experience, such as in the case of grief or intense romantic
love, or they may indicate a serious psychiatric disorder such as psychosis
(Table 15.1). They vary in the degree to which they interrupt normal thinking
and activity, whether they are perceived to originate from within the individual
or from outside, and the extent to which the individual feels they can be con-
trolled. Some intrusive thoughts may be experienced as pervasive emotional
experiences, such as worry or rumination in anxiety and depression, whereas
others may be experienced as either spontaneous intrusive experiences, such
as compulsions in obsessive-compulsive disorder (OCD). Others are elicited
by discrete environmental stimuli, such as triggers in posttraumatic stress dis-
order (PTSD) or craving elicited by cues in drug users. Intrusive thoughts also
differ in content along several modalities, including verbal, visual imagery,
and sensorimotor phenomena (e.g., urges and compulsions) and the degree to
which they can be suppressed or controlled by various behavioral procedures.
Although studying these dimensions may shed light on the psychological and
neural processes underlying intrusive thoughts, there have been few attempts
to categorize intrusive thoughts along any of these dimensions.
An obvious follow-up to recognizing this variability is to ask: To what
extent do pharmacological or behavioral treatments depend on the nature of
the intrusive thoughts? Certain drugs, for example, may be effective for intru-
sive thoughts related to depressive rumination and suicidality, whereas others
may more effectively target stimulus-elicited urges. Careful examination of
intrusive thoughts as separate entities may reveal shared mechanisms across
diagnoses, which can be revealed by focusing on medications that are effica-
cious across disorders. Naltrexone, for example, is used for opioid use as well
as for OCD, and aripiprazole is used for suicidal thoughts and for gambling.
Such analysis is consistent with the RDoC approach, designed to identify
the neurobiological and cellular mechanisms underlying psychopathological
states. Increasingly, researchers have focused on the neural features of intru-
sive thoughts. For example, Popa et al. (2016) report that stimulation in the
dorsolateral prefrontal cortex in patients with epilepsy produces the persistent
thoughts that often presage frontal seizures. Other studies (e.g., Hellerstedt
et al. 2016) have used a Think/No-Think procedure to track neural activity
related to unwanted memories using electrophysiological measures. Kühn et
Pharmacological Interventions 317
Table 15.1 Examples of intrusive events that occur in different psychiatric disorders.
Disorder Symptoms of Intrusive Thinking

OCD Recurrent and persistent thoughts, urges, or images that
are experienced as intrusive or unwanted; cause marked
anxiety and distress; individual attempts to reduce or ig-
nore thoughts and urges, or neutralize them with an action
Compulsions: repetitive behaviors that the individual feels
driven to perform
Body dysmorphic disorder Preoccupation with one or more perceived defects or
flaws in physical appearance
Eating disorders (bulimia Intense fear of gaining weight; undue influence of body
nervosa, anorexia nervosa) weight on self-evaluation
Gambling Preoccupation with gambling, persistent thoughts of reliv-
ing past gambling experiences, thinking of ways to get
money to gamble; often gambles when distressed
Pyromania Fascination with, interest in, curiosity about, or attraction
to fire; pleasure, gratification, or relief when setting fires
Kleptomania Recurrent failure to resist impulses to steal objects
Intermittent explosive Failure to control aggressive impulses (verbal or damage
disorder to property)
Major depressive disorder Rumination on feelings of worthlessness or guilt
Generalized anxiety Recurrent worries about various dimensions of life,
disorder including work, family, romance, or others
PTSD Recurrent, involuntary, and intrusive distressing memories
of a traumatic event; hyperreactivity to cues and remind-
ers of the event; avoidance of reminders of the event
Psychosis Preoccupation with delusions or false beliefs*
Substance use disorder Craving, strong desire to use a drug
*Psychosis may be part of many different disorders
al. (2013) used fMRI to show that there was greater activity in brain regions in-
volved in language production in healthy individuals who reported high, com-
pared to low, habitual tendencies for intrusive thought. This focus on intrusive
thoughts as entities in their own right provides an opportunity in the future to
investigate pharmacological interventions that target such thoughts.
Here we review the main medications that are used for psychiatric disor-
ders for which intrusive thoughts are a major feature (Table 15.2). We separate
the medications into categories based on first-line treatment, second-line treat-
ment, and experimental approaches. The medications listed under the first two
categories are considered standard care and were drawn from the online clini-
cal decision support resource site UpToDate. The medications listed under ex-
perimental approaches were drawn from a review of the literature (PubMed).
Table 15.2 Drugs used to treat psychiatric disorders with intrusive thoughts in first-line treatment, second-line treatment (where several random-
318
ized control trials have shown strong evidence), and experimental approaches (where some small studies may show effects). MAOI (monoamine
oxidase inhibitor), MDMA (3,4-methylenedioxymethamphetamine), SNRI (serotonin and norepinephrine reuptake inhibitors), SSRI (selective
serotonin reuptake inhibitors), TCA (tricyclic antidepressants).
Disorder First-Line Treatment Second-Line Treatment Experimental Approaches

OCD SSRIs (all have been FDA ap- SNRIs (e.g., venlafaxine), augmentation Psilocybin, eszopiclone, substance P, neuro-
proved except citalopram and esci- with antipsychotic (e.g., risperidone) peptide Y, vasopressin, riluzole, ketamine,
talopram), clomipramine (TCA) D-cyclosterine
Body dysmorphic disorder
SSRIs (escitalopram or fluoxetine Clomipramine Augmentation with atypical antipsychotics,
have the most evidence) oxytocin
Eating disorders
Bulimia SSRIs (e.g., fluoxetine) TCA (e.g., desipramine), trazodone,
nervosa MAOI (e.g., phenelzine), topiramate
Anorexia Psychotherapy TCA (e.g., desipramine), trazodone, D-cyclosterine plus exposure-based therapy
nervosa MAOI (e.g., phenelzine), topiramate
Gambling SSRIs Opioid antagonists Glutamatergic agents, lithium (for comorbid
bipolar disorder), topiramate
Pyromania None Mood stabilizers (e.g., lithium, carba-
mazepine, valproate, SSRIs), opioid
antagonists
Kleptomania None SSRIs, opioid antagonists
Disorder First-Line Treatment Second-Line Treatment Experimental Approaches

Intermittent SSRIs Mood stabilizer, phenytoin, oxcarbaze- Atypical antipsychotic (e.g., ziprasidone,
explosive pine, carbamazepine, lamotrigine, valpro- clozapine, risperidone, and olanzapine)
disorder ate, lithium
Major depres- SSRIs SNRIs, serotonin modulator, TCA, Ketamine, psilocybin, LSD, opioid
sive disorder MAOI, bupropion, mirtazapine medications
General- SSRIs, SNRIs Buspirone, pregabalin, benzodiazepines Eszopiclone, substance P, neuropeptide Y,
ized anxiety vasopressin, riluzole, ketamine
disorder
PTSD Psychotherapy, then SSRIs, SNRIs Second-generation antipsychotics (e.g., Mood stabilizers (e.g., tigabine, topiramate,
quetiapine), prazosin divalproex), beta-adrenergic blockers, D-
cyclosterine, ketamine, cannabis, MDMA
Substance use Opioids: Buprenorphine, Opioids: Naltrexone Multiple disorders: LSD, psilocybin, ket-
disorder methadone Cocaine: Amphetamines, modafinil, amine, MDMA, oxytocin, n-acetylcysteine
Alcohol: Naltrexone, acamprosate disulfiram, topiramate, galantamine
Tobacco: Varenicline, bupropion Alcohol: Disulfiram
Tobacco: Nortryptiline, cysticine (not
available in the U.S.)
Psychosis* Second-generation antipsychotics Lumateperone (novel antipsychotic) Cannabidiol, glutamate modulators, nicotine
(e.g., quetiapine), first-generation receptor agonists, D-cyclosterine
antipsychotic (e.g., haloperidol)
319
*Psychosis may be part of many different disorders, some of which are treated differently than the above.
In the sections below, we group the intrusive thoughts into four broad catego-
ries based on their phenomenology, known biological or environmental ori-
gins, and involvement of emotional states:
1. Behavioral: intrusive thoughts that take the form of urges and compul-
sions to engage in repetitive or unwanted actions.
2. Affective: intrusive thoughts that appear to emerge from valenced emo-
tional states, such as anxiety or depression.
3. Substance-induced: intrusive thoughts that are related to psychoactive drugs.
4. Cognitive: intrusive thoughts that involve delusional thoughts and
hallucinations.
Intrusive thoughts are rarely a primary target of treatment, and when they are,
they are usually treated with psychosocial interventions (e.g., Clark 2005;
Ainsworth et al. 2017; van Schie and Anderson 2017; Rebetez et al. 2018;
Iyadurai et al. 2019). However, as we note below, a small handful of studies
have used pharmacological techniques to reduce intrusive thoughts.
Behavioral
For several disorders, intrusive thoughts take the form of strong urges to en-
gage in harmful, apparently unnecessary, or socially unacceptable actions. This
form is a key diagnostic symptom of OCD (Clark and O’Connor 2005) and
may include thoughts about engaging in inappropriate sex acts, violence or
harm to others, or thoughts related to family members, children, or death. They
may also take the form of fear of germs or contamination as well as discomfort
at having things out of order, accompanied by urges to act (e.g., cleaning or
washing, repeated checking or repeated counting). In the case of body dysmor-
phic disorder, patients may be obsessively preoccupied with a particular aspect
of their appearance and make repeated efforts to alleviate this discomfort by
altering their bodies. Patients with OCD may worry about engaging in socially
inappropriate behaviors, such as touching or kissing someone inappropriately,
hurting someone, or engaging in actions that go against the individual’s value
system. OCD patients report escalating anxiety as they resist urges to act and a
temporary sense of relief after acting on the urge. To varying degrees, intrusive
thoughts that relate to urges to engage in socially unacceptable behaviors are
also features of other impulsive control disorders, such as eating disorders,
gambling, pyromania, kleptomania, and intermittent explosive disorder. In
each of these cases, patients become preoccupied by thoughts of engaging in
an inappropriate behavior; these thoughts are difficult to control and can inter-
fere with their normal function.
For each of these disorders, selective serotonin reuptake inhibitors (SSRIs)
are the first- or second-line of treatment. The behavioral processes through
which SSRIs relieve OCD and associated disorders have been studied in some
depth, but are still not understood. Drugs may reduce anxiety and dampen
responses to stimuli that trigger the obsessions, or they may modulate the ob-
sessive thoughts themselves. Tricyclic antidepressants are also used for OCD,
and second-line treatments include other serotonin and norepinephrine reup-
take inhibitors (SNRIs) with less serotonergic activity (Hirschtritt et al. 2017).
Cognitive behavioral therapy is an accepted nonpharmacological treatment for
OCD, and there is some support for adjuvant use of neuroleptics, which may
be especially helpful in OCD patients with tics, though may not act to reduce
obsessive thoughts (Bloch et al. 2006).
Affective
Intrusive thoughts are key features of several affective disorders, which are
characterized by strong emotional states such as depression, anxiety, or mania.
Patients with major depressive disorder report ruminative intrusive thoughts
that are congruent with their negative mood states, such as self-deprecating
thoughts of worthlessness or guilt, delusions of guilt, or paranoia in psy-
chotic depression. Patients with generalized anxiety disorder report intrusive
thoughts that are congruent with their anxious mood states, such as worries
about life, including finances, family, work performance, or accomplishing
everyday tasks. In another disorder in which anxiety is a prominent feature,
PTSD, intrusive thoughts relate directly to the experienced trauma, typically
in the form of vivid memories related to the traumatic event, including the
people, context, and emotions experienced. Individuals with PTSD exhibit a
heightened sensitivity to cues related to the trauma, which may generalize to
other stress-related cues. In PTSD, intrusive thoughts are typically associated
with strong negative affect, including both anxiety and depression. The intru-
sive thoughts themselves are highly distressing and may be accompanied by an
urge to act, in the form of fighting or fleeing from the situation. They are also
a feature of intermittent explosive disorder in which patients experience explo-
sive outbursts of anger and violence. These may be spontaneous or elicited by
inconsequential events perceived as provocation. The final category of mood
disorders, those including symptoms of mania, may be accompanied by strong
urges or recurrent thoughts that lead to impulsive behaviors, such as financial
spending or risky sexual behavior, but it is not clear to what extent these urges
are distressing to the individual experiencing them.
In major depressive disorder, intrusive thoughts which occur as part of
ruminations are typically treated with SSRIs. While ruminative thoughts are
not part of the diagnostic criteria for a major depressive episode, they fre-
quently occur as a part of the disorder. Commonly used instruments, like the
Hamilton Depression Rating Scale and the Beck Depression Inventory, do not
directly assess ruminative thinking, so little is known about the direct effect of
antidepressant medications on these types of thoughts. Future research may
utilize instruments that specifically assess ruminations, such as the response

scale Penn State Worry Questionnaire (Meyer et al. 1990), or specific ques-
tions in standardized clinical scales, such as GAD-7 (two worry questions) and
PHQ-9 (thoughts of being dead or self-harm). Beyond ruminative thoughts of
worthlessness or guilt that may present during a major depressive episode, de-
pression may also have psychotic features, characterized by delusions of guilt
and paranoia, and these require different pharmacological treatments. A recent
meta-analysis suggested that the most effective treatment for depression with
psychotic features consists of combination therapy with both an antidepressant
and antipsychotic medication (Wijkstra et al. 2013). While initially there was
a concern that novel treatments for depression, such as ketamine, may not be
appropriate for patients with psychotic features, due to potentially psychoto-
mimetic effects, case reports have suggested otherwise (Ribeiro et al. 2016).
To address the persistent worries involved in generalized anxiety disorder,
SSRIs and SNRIs are first-line treatments. Over the past several years, there
has been a shift away from treating such disorders with benzodiazepines due
to the higher risk of tolerance and higher rate of withdrawal symptoms seen
with this class of medications (Offidani et al. 2013). Other second-line treat-
ments include the azapirone buspirone and pregabalin, both of which may
act to reduce persistent worries. Buspirone may also act at 5HT-1A receptors
(Howland 2015). Pregabalin, though similar in structure to GABA, may act by
binding voltage-gated calcium channels and reducing downstream glutamater-
gic signaling (Baldwin et al. 2013). Antipsychotics have also been studied in
the treatment of generalized anxiety disorder (Gao et al. 2006). Neuropeptides,
such as vasopressin and neuropeptide Y, have shown some promise in the treat-
ment of the intrusive thoughts associated with generalized anxiety disorder,
based on their efficacy in facilitating fear extinction (Tasan et al. 2016). For
many of these treatments, it is not clear whether the medication targets the af-
fective state of fear or the anxious thoughts associated with the disorder, and
it can be hard to isolate these components in a research setting. In a promising
novel line of research, cancer patients with anxious thoughts related to their
terminal illness have been treated with single doses of psychedelic drugs, such
as psilocybin or LSD (Griffiths et al. 2016b; Ross et al. 2016). Reportedly,
80% of patients showed a reduction in anxious thoughts after only one to two
experiences with the medication. There is an urgent need for further research
to determine how single administration of psychoactive substances can induce
lasting changes in mood and psychiatric symptomatology.
In PTSD, treatments have focused on blocking different stages of the pro-
cessing of traumatic memories. Graebener et al. (2017) examined the effect
of cortisol on intrusive memories in patients with PTSD but found no effect.
Taylor and Torregrossa (2015) discuss the highly promising approach of phar-
macologically blocking the reconsolidation of maladaptive and intrusive mem-
ories, but this intriguing idea has not yet led to effective treatments. Hill et
al. (2017) discuss the possibility of targeting PTSD-related intrusive thoughts
with drugs that act on the endocannabinoid system, but such medications have
not yet reached clinical application.
Drug Induced
In substance use disorders (SUDs), intrusive thoughts are a key feature. They
typically take the form of preoccupation with drug use, reactivity to drug-related
stimuli, and craving or urges to use drugs. Indeed, “craving” was recently added
to DSM-5 (American Psychiatric Association 2013) as a symptom in the diag-
nostic criteria for SUDs, and it is among the most frequent symptom reported by
drug users attempting to quit. Cravings are problematic not only because they
predict or presage the occurrence of actual drug use, but also because they can be
sufficiently distressing to drug users to be a target of treatment in their own right,
independently of actual drug use (Green and Ray 2018). The origin of drug-
related intrusive thoughts in SUDs is not fully understood and may be multidi-
mensional, reflecting both underlying physiological conditions and responses to
Pavlovian conditioned stimuli. For example, craving is a central feature of acute
withdrawal from a drug after regular use, and it is also elicited by drug-related
cues in the environment or memories of drug use. Cravings may be elicited in
abstinent users, even long after withdrawal, by drug-related stimuli, by positive
or negative emotional events, or by ingestion of a small amount of the drug it-
self. In other cases, cravings or urges occur without any definable precipitant.
Whether all these forms of drug-related intrusive thoughts and cravings reflect a
single underlying neural process remains unknown. Notably, however, there are
intrusive thoughts specific to each drug, and withdrawal from one drug does not
induce craving for other drugs.
Although intrusive thoughts are a key feature of SUDs, they are not usually
selected as a separate target symptom for pharmacological treatment. Thus, as
with other psychiatric disorders, pharmacological treatments for SUDs target the
full constellation of the disorder rather than individual symptoms. Further, phar-
macological treatments of SUDs are typically specific for the class of drug that
is abused (e.g., opioids, nicotine). Nevertheless, several studies have examined
the effects of medications specifically on ratings of craving for a particular drug.
For example, Courtney et al. (2016) reported that naltrexone (opioid antagonist)
blocked conditioned craving to heroin cues, and Green and Ray (2018) reported
that varenicline (nicotinic partial agonist) dampened cravings for tobacco ciga-
rettes. In other studies, oxytocin blocked cravings elicited by cues related to both
cannabis (McRae-Clark et al. 2013) and tobacco cigarettes (Miller et al. 2016).
One promising new pharmacological intervention for SUDs, as well as other
psychiatric disorders characterized by intrusive thoughts, is n-acetylcysteine
(Dean et al. 2011; McClure et al. 2014; Minarini et al. 2017). N-acetylcysteine
is a precursor to the antioxidant glutathione, and it acts as a modulator of gluta-
matergic, dopaminergic, neurotropic, and inflammatory pathways. Although this
medication has not been approved for treatment, Dean et al. (2011) review prom-
ising results with this drug in the treatment of cannabis use, cigarette smoking,
cocaine addiction, and several other psychiatric conditions in which intrusive
thoughts play a role. However, larger studies on the effects of n-acetylcysteine
for cocaine addiction or cannabis addiction have been disappointing (Mardikian
et al. 2007; Gray et al. 2017).
Cognitive
Intrusive thoughts may be a part of several different psychotic disorders in

the form of delusions or cognitive distortions. Psychotic disorders, includ-
ing schizophrenia and schizoaffective disorder, may present with patients
experiencing uncontrolled and sometimes disturbing thoughts, images, or
perceptions. Delusions, or beliefs that are inconsistent with reality, are the
most obvious form of intrusive thoughts manifesting in psychotic disorders.
Perceptual hallucinations (auditory, visual, olfactory, tactile) may or may not
fall into the category of “thoughts,” depending on whether patients recognize
them as originating within their own mind. Both hallucinations and delusions
are considered “positive symptoms” of these disorders, suggesting that they
are additional thoughts and perceptions that patients struggle with, as opposed
to “negative symptoms,” which reflect a relative deficiency of volition, speech,
or other constructs.
The typical treatment of positive symptoms of psychotic disorders consist
of second-generation antipsychotic medications. Commonly used second-gen-
eration antipsychotic medications, such as olanzapine, ziprasidone, and que-
tiapine, are remarkably effective at reducing hallucinations in about 92% of
patients who take them for a year, though they have many side effects (Sommer
et al. 2012). First-generation antipsychotic medications, such as haloperidol,
are thought to target and block D2 receptors, while second-generation medica-
tions are believed to have complex actions involving D2 receptor blockade
and activity at 5HT2A receptors as well (Abi-Dargham and Laruelle 2005).
Delusions in schizophrenia and depression appear to respond to antipsychotics,
although delusions in delusional disorder and bipolar disorder are less respon-
sive to antipsychotic treatment. One study found that only 22% of psychotic
patients with schizophrenia spectrum disorders reported improvement in “cog-
nitive preoccupations” after two weeks of antipsychotic treatment (Mizrahi et
al. 2006).
New treatments for intrusive thoughts in psychotic disorders are being in-
vestigated. Cannabidiol (CBD), a component of cannabis, has been tested in
the treatment of schizophrenia. One recent study showed reductions in posi-
tive symptoms of schizophrenia and improved cognition after six weeks of
1000 mg CBD/day adjunctive treatment (McGuire et al. 2018). Glutamate
modulators have also shown promise in the treatment of intrusive thoughts in
schizophrenia (Hashimoto et al. 2013), presumably as a result of the impor-

tance of learning and the integration of new information in reducing delusional
thought patterns. Glutamate-modulator treatments are being used to facilitate
cognitive behavioral therapy, which can be used to target intrusive thoughts.
For example, the partial agonist at the glycine site of the NMDA receptor, d-
cyclosterine (50 mg), has shown promise in reducing delusional severity in
patients with schizophrenia and schizoaffective disorder (Gottlieb et al. 2011).
Sodium nitroprusside, another potential NMDA modulator, was not effec-
tive in the treatment of schizophrenia (Brown et al. 2019a). More research is
needed to determine exactly which pharmacological treatments are effective,
and in which contexts.
Neuropharmacological Considerations
It is worth considering the possibility that different neurotransmitter systems

play distinct roles in the generation of intrusive thoughts or in the ability to
control or treat them. Not surprisingly, given the heterogeneity of the intrusive
thoughts described above, almost every neurotransmitter system is implicated
in either the generation or the treatment of intrusive thoughts, including dopa-
mine, norepinephrine, serotonin, acetylcholine, glutamate, endogenous opioid,
and endocannabinoid systems, as well as hormonal systems such as oxytocin
or stress hormones. With respect to the behavioral intrusive thoughts described
above, no single neurotransmitter system has been implicated in “impulse con-
trol” disorders or OCD (Williams and Potenza 2008). Impulse control disor-
ders, such as pathological gambling, have been linked to serotonin, dopamine
(by modulating reward pathways), and norepinephrine dysfunction (arousal
and excitement). The pathophysiology of OCD is poorly understood, but sero-
tonergic drugs are commonly used to treat this disorder, and there is evidence
that glutamatergic drugs may be effective as well (Goodman et al. 2014). With
respect to affective intrusive thoughts, serotonin circuits are most strongly
implicated in both the etiology and treatment. Yet other transmitter systems,
such as GABA, norepinephrine, and most recently glutamatergic systems, are
also implicated. As with the other disorders, little is known about the recep-
tor mechanisms underlying intrusive thoughts, independent of the broader
range of symptoms of the disorders. In the case of drug-induced or substance-
abuse disorders, the focus of most treatments is on the neurotransmitter system
involved with each specific drug (e.g., cholinergic for cigarette smoking, dopa-
minergic for stimulants). Yet, there is also some evidence that certain pharma-
cological treatments may target other systems, such as oxytocin receptors or
stress hormone receptors. Other possible treatments target the reconsolidation
of drug-related memories (cues), independent of the specific drug type. With
respect to cognitive intrusive thoughts described above, the dopamine system
is most strongly implicated in the pathophysiology of psychotic disorders, and
drugs that block dopamine function are also the primary basis for pharmaco-
logical treatment. As noted above, however, the role of dopamine specifically
in intrusive thoughts, separate from other symptoms of psychotic disorders, is
not known.
Another consideration in the pharmacological treatment of intrusive
thoughts is the temporal characteristics of dosing. Many psychiatric medica-
tions, such as antidepressant and antipsychotic drugs, are prescribed for daily
use over extended periods of time. Recently there has been a growing interest
in medications that can be used with a single or small number of administra-
tions. Examples of these are the use of ketamine for depression and, more
controversially, the use of psychedelic drugs in the treatment of anxiety, de-
pression, or substance abuse. How exactly single, high doses of psychedelic or
other psychoactive drugs improve persistent psychiatric symptoms is an excit-
ing and challenging new direction of research. Do their effects depend on the
psychological experience, or are they related to neuropharmacological changes
in brain function? Finally, there is also an important role for drugs used in
combination with a behavioral intervention, such as the use of MDMA during
psychotherapy sessions or the use of beta-adrenergic blockers during presenta-
tion of drug-related memory cues. These different modes of administration of
psychiatric medications offer promising new approaches to combining phar-
macological with behavioral interventions.
Conclusions
Intrusive thoughts are symptoms of a wide range of psychiatric disorders with

markedly different pathophysiology, and they take many forms: from the urge
to act, ruminations, and cravings to delusional thoughts. Perhaps because they
are associated with so many different disorders, each with a different patho-
physiology, a wide range of medications have been used to treat intrusive
thoughts. Given the heterogeneity of their symptoms and associated disorders,
it is unlikely that any single neural circuit mediates intrusive thoughts, or that
any single medication can be used to target them across diagnostic categories.
Nevertheless, there may be commonalities associated with different disorders
(e.g., cravings of food in eating disorders or drugs in SUDs). Some intrusive
thoughts may be mainly a disorder of memory (e.g., symptoms in PTSD or
addiction), whereas others originate from disordered cognition or strong emo-
tional states. By examining intrusive thoughts both within and across disor-
ders, it may be possible to identify underlying processes that could be targets
for new pharmacological treatments.
A handful of issues warrant consideration in studying pharmacological in-
terventions for this elusive symptom. Although certain medications are clas-
sically associated with certain psychiatric disorders, in practice a wide range
of medications are used, even for the same disorder. For example, dopamine
antagonists are typically the first line of treatment for schizophrenia or sero-
tonin reuptake blockers for depression, yet physicians often sample from a
range of medications on an individualized basis to find effective treatments.
The range of pharmacological treatments is partly because patients vary in
their response to medications, as well as because the symptoms within a dis-
order vary across individuals and over time. Further, diagnoses and disorders
are in reality the composite of individual symptoms, and thus the individual
symptoms are rarely the target of medication development. It is thus difficult to
determine which medications are effective specifically for an individual symp-
tom, such as intrusive thoughts. Further, most psychiatric medications target
a constellation of symptoms involved in the disorder rather than any specific
symptom. For example, SSRIs are prescribed for depression to improve all
the components of the disorder, which may include mood, sleep, appetite, and
energy levels. This makes it difficult to examine published studies into the effi-
cacy of various medications in the treatment of intrusive thoughts. Prospective,
controlled studies with specific intrusive thought-related outcome measures
are necessary as a next step in investigating these questions.
We conclude with a comment about the recent introduction of a novel form
of pharmacological treatment for psychiatric disorders: the use of single, high
doses of “psychedelic” drugs. Drugs such as ketamine, MDMA, and psilocy-
bin, which were formerly only considered in the context of nonmedical use, are
now being tested in therapeutic settings. Remarkably, these drugs appear to be
effective after just one or two single administrations, in contrast to other medi-
cations that are used on a daily basis. It remains to be determined how single
doses of these drugs can produce lasting beneficial effects on, for example, ma-
jor depression, end-of-life anxiety, or PTSD. It also remains to be determined
whether, and how, these novel treatments affect intrusive thoughts in particular.
This is an exciting and promising new direction for psychiatric medications.
Acknowledgments
HdW was supported in part by NIDA DA02812 and AKB was supported by a training
grant from the National Institute of General Medical Sciences (2T32GM007281). We
thank Larry Price for helpful comments on the manuscript.
16
Developing a Neuromodulation
Tool to Suppress
Intrusive Thinking
Things We (Think We) Know and
Things We Need to Figure Out
Colleen A. Hanlon and Lisa M. McTeague
Abstract
Intrusive thinking is a core feature in multiple psychiatric diseases, including obsessive-
compulsive disorder (OCD), posttraumatic stress disorder (PTSD), substance use dis-
order (SUD), and Tourette syndrome. These diseases are not only bound by intrusive
thinking, they also share similar disruptions in the functional architecture of the brain,
including frontal-striatal-thalamic circuitry which is involved in salience attribution and
shifting attention. As more is learned about the neural circuit dysfunctions involved in
the initiation, maintenance, and attention to intrusive thoughts, it may become possible
to develop noninvasive neuromodulation approaches to attenuate the presence of these
thoughts or the morbidity associated with their existence in individuals. This chapter
focuses on transcranial magnetic stimulation (TMS) as a tool to induce causal change
in behavior, cortical excitability, and frontal-striatal activity. An overview is provided
of the cortical and subcortical areas that are often implicated in intrusive thinking, using
examples from Tourette syndrome, OCD, PTSD, and SUD. The hypotheses presented
can be generalized past TMS to other invasive and noninvasive forms of neuromodula-
tion. In conclusion, key questions are posed to move the field forward.
Introduction
As discussed in the other chapters of this book, there are many operational
definitions for intrusive thoughts. While it is difficult to unify these defini-
tions into one common framework, a familiar theme is present in all of them:
330 C. A. Hanlon and L. M. McTeague
a persistent, stereotyped mental pattern that aversively interrupts the flow of

competing mental processes despite attempts to inhibit and/or counter these
thoughts. One particularly compelling example is that intrusive thoughts are
like a mental “tic” wherein, as in the case of Tourette syndrome, an individual
may be able to suppress intrusive thoughts for some period of time, but even-
tually that cognitive buffer is broken down and the intrusive thought pattern
floods the neural systems that kept it in check.
The clinical disorder that is most easily characterized as an impairment
of intrusive thinking is obsessive-compulsive disorder (OCD), wherein re-
current thoughts or urges lead to debilitating levels of anxiety, distress, and
resultant compulsive actions that patients typically fail to willfully suppress.
The anxiety and distress associated with intrusive, unwanted thoughts are
also a hallmark of posttraumatic stress disorder (PTSD). Similarly, intrusive
thoughts about avoiding opiate or alcohol withdrawal or having time to “take
the edge off” with a smoking break, fuel the growth of substance use disorder
(SUD) and impair the ability for treatment-seekers to remain abstinent.
All four of these psychiatric conditions (Tourette syndrome, OCD, PTSD,
and SUD) have intrusive thoughts at their core, yet existing behavioral and
pharmacological treatment strategies for these diseases are very different.
Modern psychiatry has only recently begun to approach disease treatment in a
manner that focuses on core transdiagnostic symptoms of psychiatric disease
rather than discrete disease labels. Inasmuch as intrusive thinking is a core
symptom common to these disorders, it is certainly a research domain worthy
of focus for treatment development.
In addition, these four diseases have something else in common: they all
entail disruptions in frontal-striatal circuitry involved in limbic drive and cog-
nitive control. Through recent advances in dosing and coil design, it appears
that a noninvasive brain stimulation technique known as transcranial magnetic
stimulation (TMS), approved by the U.S. Food and Drug Administration (FDA)
to treat depression and OCD, may be a promising tool to target the functional
neurocircuit substrates of intrusive thinking in these patients. TMS is one of
several noninvasive techniques that can be used to modulate neural circuitry
(see Figure 16.1). Here we introduce TMS as a tool to induce causal change in
behavior, cortical excitability, and frontal-striatal activity. We provide an over-
view of the cortical and subcortical areas that are often implicated in intrusive
thinking (using examples from Tourette, OCD, PTSD, and SUD) and outline
several key questions that should be addressed to move the field forward.
The Application of Transcranial Magnetic

Stimulation to Diseases of Intrusive Thinking
TMS is a noninvasive brain stimulation technique that can induce changes in

neural activity in the cortex and in monosynaptic afferent projections. When
Developing a Neuromodulation Tool 331
Conventional TMS Deep TMS tES (tDCS/tACS)

Figure 16.1 Noninvasive neuromodulation techniques used in individuals with
psychiatric disorders that involve intrusive thoughts. The most common technique is
the conventional transcranial magnetic stimulation (TMS), which is done by placing
a figure-of-eight coil over a specific cortical location. This technique has been used
to modulate craving in substance use disorder, impulse control in Tourette syndrome,
obsessions in obsessive-compulsive disorder (OCD), and general symptoms of post-
traumatic stress disorder. A unique form of TMS, known as “deep TMS,” uses similar
technology to modulate a wider, deeper area of cortex. This technique was approved
for treatment of OCD in 2018. Transcranial electrical stimulation (tES) includes tran-
scranial direct current (tDCS) and alternating current (tACS) approaches and has been
used in these disorders as well, although there is still not clear evidence of its clinical
efficacy. Reprinted with permission from Ekhtiari et al. (2019).
delivered repetitively (e.g., 600–3000 pulses every 40 seconds to 20 minutes),

TMS can change cortical excitability and various behavioral phenomena for
30 minutes to several hours. When these repetitive sessions are given sequen-
tially over a series of days (e.g., 10–30 sessions over 2–6 weeks), there may
be lasting changes in functional connectivity in the brain as well as behavioral
symptom resolution for several months to a year.
TMS was approved by the FDA as a treatment for major depressive dis-
orders in 2008, and there are now TMS clinics in all 50 states in the United
States, throughout Europe, Asia, Australia, and South America as well as a
few new clinics in Africa. While the majority of the research in TMS has fo-
cused on optimizing treatment protocols for depression, there has been an ex-
ponential growth in the application of TMS to investigate and modulate these
networks in populations with Tourette syndrome, OCD, PTSD, and SUD. The
data has been growing fast, such that in 2018 the FDA approved a unique form
of TMS to treat OCD. There is already approval for its use as a tool in SUD
and OCD in Europe.
In developing a noninvasive neuromodulation solution for intrusive think-
ing, however, many questions remain:
• Is there a common neural circuitry that drives intrusive thinking across
disease states? If so, can we use the same stimulation protocol for ev-
eryone or will there be biotypes that should be considered?
• Assuming that we focus on TMS (as it is the only FDA-approved, non-

invasive neurostimulation technique), what is the best cortical location
and frequency?
• What is the best way to combine neurostimulation with pharmaco-
therapy and behavioral therapy to attenuate intrusive thoughts? Should
these techniques be given simultaneously or in serial?
• Should we be pursuing closed-loop neuromodulation strategies for in-
trusive thinking (rather than current open-loop approaches)? Can we do
that noninvasively?
• What stage of intrusive thinking is the optimal target for remediating
intrusive thinking? Should the focus be on prevention (i.e., prevent
the initiation or exacerbation of intrusive thoughts), inhibition (i.e.,
suppress intrusive thoughts), reframing (i.e., change the valence of a
positive/negative thought), or distraction (i.e., enable a patient to shift
attention away from the thoughts)?
Below, we will attempt to provide some insight into these questions. As a basis
for this discussion, we begin with a review of several key principles that are
important to understand, in terms of the capabilities and restrictions of current-
generation TMS devices.
What Is Transcranial Magnetic Stimulation?
TMS can modulate neural excitability. It is a noninvasive form of brain stimu-

lation that induces a depolarization of neurons through electromagnetic in-
duction. Although a comprehensive review of studies that have demonstrated
the principles of TMS is beyond the scope of this chapter, prior behavioral,
electrophysiological, and neuroimaging work in this area is well described
and summarized in several review articles (Fitzgerald and Daskalakis 2008;
Hoogendam et al. 2010). The majority of our knowledge regarding the basic
electrophysiological effects of TMS on the brain are from studies in the mo-
tor system. When applied over the hand knob of the primary motor cortex, a
single, transient pulse of current through the TMS coil induces a reliable con-
traction of the contralateral hand, proportional to the amplitude of the induced
electrical field (Barker et al. 1986). The amplitude of this motor-evoked po-
tential (MEP) in the contralateral hand can be manipulated by pharmaceutical
agents that effect voltage-gated sodium channels and glutamate (Ziemann and
Rothwell 2000; Di Lazzaro et al. 2008). There is a dose-response relationship
between the amplitude of the TMS pulse and the amplitude of the MEP. This
dose-response relationship is referred to as the “recruitment curve” in brain
stimulation literature and can be used as a measure of cortical excitability.
TMS can modulate neural pharmacology. Although hundreds of studies
have evaluated the effects of various repetitive TMS protocols on behavior and
cortical excitability (via EEG and functional MRI), very little is known about
the effects of rTMS on neuropharmacology. The most cited studies in this do-
main have been done using positron emission tomography (PET) imaging,
wherein the radioligand is given to the participant and then the TMS stimula-
tion is delivered before the participant goes into the PET scanner. Using PET
imaging, Strafella et al. (2001) have demonstrated that dorsolateral prefrontal
cortex (dlPFC) stimulation leads to an increase in dopamine binding in the
caudate. They also showed that when 10 rTMS is delivered to the left primary
motor cortex, increases in dopamine are seen in the ipsilateral, left putamen.
Using magnetic resonance spectroscopy, several studies have demonstrated the
effects of TMS on cortical γ-aminobutyric acid (GABA) and glutamate (Stagg
et al. 2009; Vidal-Pineiro et al. 2015; Iwabuchi et al. 2017). One of the most
cited studies, by Stagg et al (2009), demonstrated that the attenuating effects
of inhibitory, continuous theta burst stimulation (cTBS) on cortical excitabil-
ity are related to an increase in GABA at the area of stimulation rather than
a change in glutamate. Recently, Iwabuchi et al. (2017) showed that a single
session of excitatory, intermittent theta burst stimulation (iTBS) to the dlPFC
leads to a decrease in the GABA/glutamate ratio in both the dlPFC and in the
insula, suggesting that it is possible to modulate paralimbic cortex through
superficial PFC stimulation.
The following principles need to be considered as TMS therapeutic strate-
gies are developed to address intrusive thinking.
Principle 1: Stimulation Depth
With a growing number of TMS coil designs, the depth at which stimulation
should occur has become increasingly complex. The focality of TMS is related
to the shape of the coil. There is a substantial body of literature devoted to
computational modeling of electric field distributions associated with differ-
ent coil shapes. In one of the most comprehensive papers, Deng et al. (2013)
investigated the focality and penetration depth of 50 existing TMS coils.
Their computational models revealed that typical figure-of-eight coil designs
affected approximately 10 cm2 of cortical surface, circular coils affected ap-
proximately 50 cm2, and H-coil designs affected approximately 100 cm2. Most
flat figure-of-eight and circular coil designs had penetration depths of 1–2 cm2,
whereas the H-coil designs had consistently higher depths of 2–3 cm2. A single
TMS pulse from a standard figure-of-eight coil stimulates a 12.5 cm2 area,
which is approximately 1/125 (0.8%) of the cortical surface area. By compari-
son, deep brain stimulation can be at least an order of magnitude more precise
than the most focal TMS coils available, with stimulation volumes ranging
from 1–20 cm, depending on the electrode configuration (Wei and Grill 2005).
Electroconvulsive therapy, on the other hand, appears to effect 94% of the
brain and magnetic seizure therapy effects 21% of the brain (Lee et al. 2016).
To put the focality of TMS in context with something that is meaningful to the
average curious member of the public, 1/125th of the cortical surface is roughly
analogous to the surface area of India (or half of Australia) relative to Earth
(Hanlon 2017).
Principle 2: Polysynaptic Transmission
Beyond the direct cortical effects of TMS, it is possible to modulate monosyn-

aptic (and possibly polysynaptic) targets of these cortical areas (Figure 16.2).
The indirect effects of cortical TMS on monosynaptic afferent targets can be
demonstrated through a behavioral assessment of the recruitment curve. When
TMS is applied to the hand area of the primary motor cortex there is a dose-
dependent change in the MEP of the hand contralateral to the TMS coil. This
pathway from the motor cortex to the hand requires at least two neurons: the
upper motor neuron, which originates in the motor cortex and terminates in
the spinal cord, and the lower motor neuron, which originates in the spinal
cord and terminates in the muscles that will contract to produce the MEP. The
majority of upper motor neurons, however, terminate on interneurons, which
then facilitate lower motor neuron activity. This suggests that TMS can have
polysynaptic effects.
(a) (b) (c)

BOLD response to BOLD response to
left dlPFC TMS mPFC/frontal pole TMS
R L
Figure 16.2 Polysynaptic effects: (a) Single pulses of TMS delivered to the hand
knob of the primary motor cortex are able to transmit information down the cortical
spinal tract, which crosses the synapse in the ventral horn, leading to contraction of
the efferent target muscle in the hand, measured with motor-evoked potentials. (b) This
polysynaptic engagement can be demonstrated in the cortex as well, wherein single
pulses of TMS delivered to the left dorsolateral prefrontal cortex (dlPFC) lead to el-
evated BOLD signal in the dorsal striatum and ventral cingulate, whereas (c) TMS to
the left frontal pole leads to BOLD signal in the ventral striatum, dorsal cingulate, and
anterior insula. Adapted after Hanlon (2017).
Principle 3: Frequency-Dependent Modulation
As stated above, when single pulses of TMS are delivered in rapid succession
(rTMS), it is possible to change cortical excitability and various behavioral
phenomena for a relatively brief period of time (e.g., 30 minutes to several
hours; see Figure 16.3). These effects appear to be frequency dependent: low-
frequency, continuous stimulation decreases cortical excitability whereas high-
frequency, intermittent stimulation leads to an increase in cortical excitability
(reviewed in Fitzgerald et al. 2006; Thickbroom 2007).
One of the first studies in this field was conducted by Pascual-Leone et al.
(1994), who discovered that 20 pulses at 10 Hz and 20 Hz stimulation over the
motor cortex produced an increase in the amplitude of the MEP, suggesting this
frequency increases cortical excitability. Chen et al. (1997) then demonstrated
that 15 minutes of 0.9 Hz TMS stimulation (810 pulses) to the motor cortex
would decrease motor cortex excitability. In a sample of 14 individuals, 1 Hz
TMS to the motor cortex for 15 minutes decreased the MEP by 20% for at
least 15 minutes after stimulation. These data are compatible with preclinical
electrophysiology studies which have demonstrated that 1 Hz stimulation in-
duces long-term depression of neural activity in slice preparations of the motor
cortex, visual cortex, and hippocampus.
While 10 Hz and 1 Hz TMS are still widely used, a unique bursting fre-
quency known as human theta burst stimulation (TBS) has now gained sig-
nificant traction in the field. Human TBS was first evaluated by Huang et al.
(2005). Leveraging data from preclinical literature, which demonstrated that
Standard repetitive TMS: ~3000 pulses Effect on cortical

excitability
1s
1 Hz
10 s
10 Hz
5s
Theta burst TMS: 600 pulses (50 Hz triplets x 5/sec)

Continuous
TMS
8s
Intermittent
TMS
2s
Figure 16.3 Frequency-dependent effects: When delivered in a repetitive manner, a

single session can effect cortical excitability for 30–60 minutes.
electrical stimulation of cortical slices in 100 Hz bursts five times per sec-
ond (known as theta burst) can induce long-term plasticity (Bear and Malenka
1994; Malenka and Bear 2004), Huang et al. performed a clinical TMS study
wherein TMS pulses were delivered to the motor cortex in 50 Hz bursts five
times per second (human TBS). When TBS was delivered continuously for
600 total pulses, it decreased motor cortex excitability. When TBS was de-
livered in an intermittent pattern (2 sec on, 8 sec off) for 600 pulses, excit-
ability increased. The effect sizes of these brief continuous TBS (40 sec) and
intermittent TBS (190 sec) paradigms are comparable to studies of 1 Hz and
10 Hz. However, many publications have recently shown that there is high in-
terindividual variability in TBS response, which has led to some caution in the
reliance on this stimulation protocol (Vernet et al. 2014; Jannati et al. 2017).
Principle 4: Priming and State-Dependent Effects
A large body of literature demonstrates that the effects of TMS on behavior are
brain-state dependent and may be amplified by priming the brain with either a
behavioral task or brain stimulation (Opie and Cirillo 2017). One of the earli-
est studies in this field was by Iyer et al. (2003), who demonstrated that the
attenuation of cortical activity with 1 Hz TMS can be amplified by priming
the motor cortex with 6 Hz TMS. This was expanded to studies in the motor
system and visual system, which demonstrated that there were brain-state de-
pendent effects of TMS on cortical excitability (Silvanto et al. 2007, 2008a, b).
Additionally, priming the brain with continuous TBS may enhance efficacy of
intermittent TBS (Opie et al. 2017).
Although this body of research existed in sensory and motor control litera-
ture, it has only recently been harnessed by the clinical TMS research field.
Whereas the recent FDA approval of TMS for OCD requires a behavioral
prime, for example, neither the brain state nor the behavioral state of the indi-
vidual was accounted for during the initial multicenter clinical trials of TMS
for depression. This represents a latent opportunity for us to improve outcomes
and minimize some of the interindividual variability that is observed in pa-
tients receiving clinical TMS treatment.
Within the addiction literature, a large clinical trial demonstrated that ex-
posing a smoker to smoking cues (behavioral prime) before TMS amplified
the effects of TMS on smoking cessation (Dinur-Klein et al. 2014). In this pro-
spective, double-blind, sham-controlled study, 115 regular cigarette smokers
were randomized to receive ten daily treatments of TMS. Immediately before
each session, half of the participants were presented with visual smoking cues:
cigarette consumption and nicotine dependence were reduced, and the effects
were greatest in individuals that were exposed to smoking cues. In PTSD treat-
ment, priming a trauma memory at the outset of each rTMS session has also
been shown to enhance TMS effect sizes (Isserles et al. 2013). In this study,
thirty PTSD patients were randomized to one of three groups: sham rTMS,
real rTMS following exposure to a 30-second patient-tailored trauma script,

or real rTMS following exposure to a 30-second patient-tailored neutral script.
Participants received 12 sessions of real or sham rTMS (three sessions per
week for four weeks, deep TMS H-coil). The only group with a significant
improvement in the Clinician-Administered PTSD Scale was the group that
received exposure to trauma scripts before each rTMS treatment.
The aforementioned studies all demonstrate that a priming stimulus ampli-
fies the effects of a form of TMS intended to increase excitability in targeted
networks. Although the mechanism through which cue exposure enhances the
behavioral effects of rTMS are not clear, one possibility is that cue exposure
reactivates latent memory traces, frequently referred to as an engram (Vernet
et al. 2014), enabling them to be manipulated and reconsolidated (Opie et al.
2017). If this were true, priming may also be effective for TMS paradigms
designed to decrease cortical excitability (e.g., 1Hz, cTBS). A recent study
by our group demonstrated that, on average, individuals with cocaine use dis-
order, who were exposed to cocaine cues before and after continuous TBS,
had a decrease in cocaine-cue reactivity following 3600 pulses of real but not
sham TBS (Malenka and Bear 2004). Secondary analyses of the data, however,
demonstrated that real cTBS decreased cue reactivity in individuals with a high
baseline brain response to cues, whereas it increased cue reactivity in individu-
als with a low baseline brain response to cues. This bidirectional shift was not
present following sham cTBS. While all individuals in this study received a
behavioral prime (drug cue exposure), it seems that a behavioral prime was not
alone sufficient. The directionality of TBS-induced effects was dependent on
the baseline level of brain activity (neural state) in the TBS target; in this study
(Malenka and Bear 2004), the medial PFC.
Potential Neural Circuit Targets for

Neuromodulation of Intrusive Thinking
One of the key advances in the neuroimaging literature over the last twenty
years is that brain regions organize their activity into coherent functional net-
works (Figure 16.4). Through functional magnetic resonance imaging (fMRI),
these networks appear as correlations of the low-frequency fluctuations in
BOLD signal between brain regions. Many networks were originally identified
via data-driven methods from brain activity at rest, and are called resting-state
networks. However, these networks reliably appear in ongoing brain activity
during tasks, and meta-analyses of task-based activation also reveal consistent
functional networks similar to those identified at rest.
Several functional networks have been studied extensively that have rele-
vance to intrusive thinking: the default mode network (DMN), containing the
medial PFC and posterior cingulate; the salience network (SN), containing
the anterior cingulate and anterior insula; and the executive control network
pre-SMA/dACC
(OCD target)
dlPFC
(executive control
mPFC network node)
(default mode
network node)
Thalamus Striatum
Globus Pallidus Interna
Ventral
Figure 16.4 Candidate neural circuits amenable to modulation for intrusive thinking.
Typically, therapeutic neuromodulation approaches require that a specific neural system
be identified. The therapeutic strategy can then target the nodes of this neural system.
This can be done invasively (e.g., through deep brain stimulation) or noninvasively
(e.g., using TMS, transcranial electrical stimulation, pulsed ultrasound). Three corti-
cal nodes may be putative targets for neuromodulation based on their role in intrusive
thinking and their striatal-thalamic connectivity: the medial prefrontal cortex (mPFC,
red), the left and right dorsolateral prefrontal cortex (dlPFC, yellow), and the presupple-
mentary motor area (pre-SMA, blue). The striatal, pallidal, and thalamic nodes of these
circuits are shown in the lower panels. Adapted after Morris et al. (2016).
(ECN), containing the dlPFC and posterior parietal cortex. DMN is the best
known and most studied of these functional networks. It serves various in-
trospective functions related to intrusive thinking, including mind wander-
ing, recollection and prospection, rumination, and self-reflection. Other “task
positive” networks act in opposition to the DMN. These networks activate
during behaviorally regulated task performance and externally focused cog-
nition. For example, the SN activates for transitions from introspection to
task performance as well as during task initiation and switching. The ECN
is involved in cognitive control, working memory, and in tasks governed by
external stimuli, whereas functional connectivity in the DMN is typically
high during tasks of internal monitoring. In this manner, the ECN and DMN
are considered anticorrelated networks.
Etkin and colleagues demonstrated the central importance of the SN as a
common neural substrate across psychiatric illness categories (Goodkind et al.
2015). They performed a meta-analysis of structural abnormalities across six

psychiatric disorder categories, including OCD, PTSD, and SUD, and found
that all of them showed gray matter volume reductions in the dorsal anterior
cingulate cortex and anterior insula. In a parallel meta-analysis of functional
neurocircuit anomalies during cognitive processing tasks across psychiatric
disorders, Etkin and colleagues demonstrated that the SN, in conjunction with
the broader frontoparietal ECN, is hyporeactive among these patients during
cognitive demands. Importantly, the ECN is recruited for cognitive regulation
as well as emotional regulation, and the dlPFC target most frequently utilized
in therapeutic rTMS is seated within this network. As such, the transdiagnostic
effects of rTMS may, in part, be attributable to ECN upregulation and its influ-
ence on attenuating intrusive thinking and associated negative affect. In addi-
tion to putatively increasing activity of the ECN with rTMS, we might predict
that intrusive thinking could alternatively be attenuated by either decreasing
the activity of the DMN or enabling the SN to switch more effectively from
the DMN to the ECN. These hypotheses have not been directly tested but are
amenable to systematic evaluation through TMS.
Questions and Hypotheses
1. Using TMS (because it is the only FDA-approved noninvasive technique),

what is the best cortical location and frequency?
Strategy 1 Decrease the amplitude of intrusive thoughts at rest by dampen-

ing the medial orbitofrontal cortex (OFC) connectivity (a node in the DMN).
Several studies have demonstrated that individuals with OCD have elevated
activity in medial aspects of the OFC. Three studies that targeted the OFC
with TMS showed improvements in OCD symptoms. These were all relatively
small studies that applied 1–3 weeks of treatment and showed changes that
lasted up to one month.
Strategy 2 Increase control over intrusive thoughts by increasing dlPFC con-

nectivity (a node of the ECN). The dlPFC is the FDA-approved target for the
treatment of depression and has been investigated as a treatment target for
PTSD, SUD, and OCD. The first study to explore the use of TMS for OCD
demonstrated that a single session of 10 Hz TMS decreased compulsions
but not obsessions, yet the effects lasted for eight hours (n = 12 individuals).
Although these results were promising, they have been difficult to replicate.
Two recent studies have demonstrated that several weeks of TMS may im-
prove OCD, but again, the obsessive component does not seem to respond
very well. This may be because the obsessive component relies more heavily
on subcortical structures such as the basal ganglia and amygdala. While 10 Hz
TMS to the left dlPFC and 1 Hz TMS to the right dlPFC have been evaluated
as a target for PTSD, with mixed success, intrusive thinking is not often re-
ported as a primary outcome measure. Consequently, with mixed results from
the dlPFC, it is still unclear if targeting this area is likely to improve intrusive
thoughts.
Strategy 3 Target the supplementary motor area/pre-SMA. The first studies

that targeted the pre-SMA examined the use of TMS for patients with OCD and
Tourette syndrome. At the end of treatment, patients showed general reduction
in OCD symptoms, an improvement in functioning, and reductions in depres-
sion and anxiety. Importantly, the improvements held for at least three months.
This study was followed by a second study with 21 OCD patients and a more
careful study design. After four weeks of TMS treatment, patients showed no-
table decreases in OCD symptoms as well as a reduction in depression and
anxiety; benefits were still present for most patients three months later.
Although TMS targeting the pre-SMA has been shown to be the most effec-
tive, it is not clear whether this is indeed the only or best area of the brain to
target, as pre-SMA studies are the only ones thus far that use doses and treat-
ment protocols similar to the standard of care for depression. These positive
results are, however, very encouraging and are helping us move forward.
In 2018, a unique form of deep TMS was approved by the FDA to treat OCD.
This type of TMS has a wider cortical field and likely modulates a large portion
of the medial PFC and cingulate cortex, including the SMA/pre-SMA. Although
it is not clear exactly which of these brain regions is responsible for the clinical
effect (or if all are necessary), these data suggest that the dorsal medial wall of the
PFC may be a good target for modulation. In an interim analysis of a larger study,
the research team evaluated the effects of 10 Hz, 1 Hz, and sham TMS on OCD
symptoms in 23 individuals (25 sessions over five weeks). They demonstrated
that although there was no significant interaction between group and time with
this sample size, the effect size was higher with 10 Hz TMS compared to 1 Hz
TMS. Hence, the remainder of the participants were randomized to 10 Hz TMS
or sham TMS for a total sample of 30, wherein 10 Hz led to a significant reduc-
tion in OCD symptoms up to one month after the five weeks of TMS (Carmi et
al. 2018). These data led to an 11-site clinical trial of 42 individuals who received
six sessions of daily TMS to this medial PFC target.
2. What is the best way to combine behavioral therapy (including mind-
fulness and neurofeedback) with TMS to maximally attenuate intrusive
thoughts? Should this be given simultaneously or in serial?
As described earlier, there is growing evidence that the effects of TMS can
be amplified by priming the individual (and perhaps pushing the brain into
a specific state) before TMS is administered. One of the key components of
the 2018 FDA approval of TMS for OCD was that TMS had to be given in
the presence of a personalized visual cue that caused stress and anxiety for
the patient (e.g., placing a purse on the dirty floor in front of someone with
obsessional thoughts about dirt). It was assumed that this external cue places
the brain in a primed state, which would then be modulated by the TMS.
Based on these empirical results from rTMS experiments and from a strong
preclinical foundation regarding manipulation and reconsolidation of mem-
ories, it is likely that any neuromodulation approach for intrusive thinking
should involve putting the individual in a state where the intrusive thoughts
are present and perhaps bothersome.
3. Should we pursue closed-loop neuromodulation strategies for intrusive

thinking (rather than current open-loop technologies)? Can we do this
noninvasively?
Given that intrusive thoughts are frequently transient and share some of the
same temporal properties of a seizure (e.g., largely unpredictable onset time
but with some known triggers, an episodic disease feature rather than a stable
state as is seen in mood disorders or chronic pain), a closed-loop system may
be more effective and appropriate for intrusive thinking than an open-loop neu-
romodulation approach (Figure 16.5).
Open-loop neuromodulation typically refers to a device that provides a
fixed stimulation protocol over a fixed period of time. Currently, most invasive
and noninvasive neuromodulation approaches are open loop (e.g., TMS, DBS).
It is easy to see the value of a closed-loop system, which could include “sens-
ing technology” to detect changes in the brain state, and then dynamic stim-
ulation settings, which could adjust to the individual patient’s neural needs.
Closed-loop stimulation technology has shown promise as a treatment for vari-
ous diseases (Widge et al. 2017). The most successful closed-loop system in
clinical trials thus far has been developed, FDA approved, and is now deployed
for use in intractable epilepsy. The leads of this device (NeuroPace) are im-
planted into an epileptic focus in the brain. It monitors activity in the areas,
can detect the prodrome of seizure activity, and when a seizure begins it will
deliver stimulation to block the growth of that activity. This is referred to as re-
sponsive neurostimulation and received approval from the FDA in November,
2013. In the years since its approval, it has been successfully employed in
hospitals throughout the United States: several trials demonstrated a 53% sei-
zure reduction after two years and a 70% median seizure reduction after five
years (Geller et al. 2017). Closed-loop stimulation has also been used in the
spinal cord for pain management (the RestoreSensor™ System, Medtronic,
Minneapolis, MN). Although no noninvasive, closed-loop brain stimulation
devices have been approved for clinical use yet, there is extensive growth in
this field (Bergmann et al. 2016).
There are also non-device-based closed-loop neuromodulation strategies.
Simpler versions include a “human in the loop”: clinicians observe the re-
corded brain signals and provide manual adjustment of the stimulation rather
than the device automatically self-adjusting. Other techniques such as real-time
342
Open-Loop Stimulation Closed-Loop Stimulation Developing closed-loop neuromodulation

(in practice) (in practice) for intrusive thinking
Calibration session Tools for sensing:

Calibration session (done manually by • Neural firing (invasive)
(done manually by including a biomarker) • EEG, rtfMRI? (noninvasive)
clinician) • Heart/pulse rate (noninvasive)
• Patient perception
Device monitors neural Patient returns to
activity via sensing clinician Transmission: wires or wireless
Patient technology
STIMULATION as returns to
PROGRAMMED clinician Tools for stimulating:
STIMULATION as PROGRAMMED if • DBS (invasive)
predefined biomarker level (or • TMS,TDCS, ultrasound (noninvasive)
behavioral sympton) is detected
Figure 16.5 Open- and closed-loop neuromodulation options for intrusive thinking. Given the temporal profile of intrusive thinking, a closed-
loop neuromodulation approach may be more beneficial—one that includes the ability to detect aberrant brain states (or behavioral states) and
adapt its stimulation properties accordingly (e.g., turning itself on/off, modifying the frequency, changing the amplitude). A critical element in-
volved in closed-loop technologies, however, is the reliability of the underlying biomarker it is trying to sense. In practice, standard “open-loop”
techniques such as TMS and DBS are actually closed-loop systems, because the patient always returns to the clinician to have the parameters ad-
justed. New generation closed-loop techniques, however, may enable these adjustments to be made in real time, responding to endogenous neural
changes and increasing the “agency” of patients by enabling them to modify the settings based on their own perceptions. This may be particularly
useful for intrusive thinking.
fMRI feedback, EEG feedback, and mindfulness strategies may also be useful
approaches for intrusive thinking.
4. What is the best treatment strategy: preventing the initiation of intrusive
thoughts or increasing the ability of the patient to shift their attention away
from the thoughts?
When designing a neuromodulation approach for intrusive thoughts, the
neural regions targeted and the dosing parameters used will likely be differ-
ent if the goal is to help individuals shift their attention away from intrusive
thoughts than if the goal is to stop them from happening. If the goal is to
stop the initiation of intrusive thoughts, it is possible that targeting the DMN
(perhaps using TMS directed at the medial PFC) might be a fruitful strategy.
Alternately, if the goal is to enable individuals to shift their attention away
from thoughts which target the salience network, a deep form of TMS di-
rected at the cingulate or insula might be the best strategy, given its role in
set-shifting and attributing value.
Conclusion
Intrusive thoughts are a common, transdiagnostically relevant feature of many

psychiatric conditions including Tourette syndrome, SUD, PTSD, and OCD.
With the approval of TMS as a tool to treat OCD in 2018, we are in the early
stages of an era of rapid discovery regarding the use of neuromodulation to
alter intrusive thoughts that plague these patient populations. Although several
concepts of rTMS treatment are robust and replicable (e.g., regional specificity,
depth of the magnetic field, dose-dependent amplification of behavior, poly-
synaptic engagement, frequency-dependent effects), many key components of
TMS treatment development have not yet been widely explored, especially for
intrusive thinking (e.g., optimal number of sessions per day or in total, the use
of behavioral primes to amplify TMS treatment effects, the effects of applying
TMS before vs. after behavioral therapy, the use of TMS to amplify pharmaco-
therapy treatment). As study of the neural circuitry that underlies the initiation,
maintenance, and distraction from intrusive thinking matures, we will be better
prepared to design biologically informed and rigorous neuromodulation clini-
cal trials in this domain.
Here, we have attempted to introduce TMS as an innovative new tool which
can modulate brain activity in a circuit-specific, frequency-dependent man-
ner as well as to review current knowledge regarding the pharmacologic ef-
fects of TMS. While development of TMS as a new treatment tool is still in
its infancy, we hope to have sparked interest in the need to develop a neural
circuit-based treatment tool—one that is available to our patients—and to in-
crease our knowledge of the synergy between pharmacotherapeutics and brain
stimulation interventions. A large body of knowledge suggests that frontos-

triatal circuit activity is a significant biomarker involved in intrusive think-
ing. Through closed-loop brain stimulation techniques, it may be possible to
develop an adaptive, personalized neural circuit-based treatment for patients.
It is important to note that lasting behavioral change may require more than
just brain stimulation. Just as the plasticity potential of a primed neuron is
higher than an unprimed neuron, TMS may have higher efficacy when an in-
dividual is engaged in the cognitive/emotional process they wish to amplify
or attenuate. Hence, TMS is likely to be most effective when combined with a
pharmacotherapeutic agent that lowers the threshold for cortical excitability or
with behavioral interventions (e.g., exposure therapy or contingency manage-
ment). Nonetheless, while these statements are based on preclinical literature
and human learning theory, they await rigorous evaluation.
As the field continues to grow, we hope to see more interactions between
clinical and preclinical neuroscience researchers from electrophysiological
and pharmacological backgrounds. With any luck, through the continued re-
finement of open- and closed-loop brain stimulation tools, we may soon be
rigorously evaluating noninvasive brain stimulation solutions for intrusive
thinking. The quest for a sustainable treatment solution will undoubtedly re-
quire a complementary approach to modifying the pharmacology, neural cir-
cuitry, and ultimately the behavioral manifestations of intrusive thinking in
these complementary cohorts of patients.
Acknowledgments
Figure 16.4, from Morris et al. (2016), is used under the terms of the Creative Commons
Attribution 4.0 International License (CC-BY).
17
Interventions and Implications

Judson Brewer, Harriet de Wit, Aurelio Cortese,
Damiaan Denys, Colleen A. Hanlon, Emily A. Holmes,
Martin P. Paulus, Jens Schwarzbach, and Peter Tse
Abstract
This chapter provides a framework for developing interventions that specifically

target intrusive events. It describes the challenges in defining intrusive thoughts
and the difficulty in distinguishing normal processes of cognition and emotion from
indicators of dysfunction, defined from practical, neurobiological, or cultural points
of view. Throughout, the term intrusive events is used to encompass both thoughts
and images that become intrusive. Examples are explored as they occur in different
psychiatric disorders to demonstrate their variance in form, frequency, and control-
lability. Treatment modalities that have been used to alleviate intrusive events in
different psychiatric disorders are reviewed, including behavioral, pharmacologi-
cal, and emerging electromagnetic brain interventions. Two clinical vignettes il-
lustrate the nature and severity of intrusive events in patient populations as well
as the complex, multidimensional nature of the clinical reality. Ways of measuring
intrusive events are examined and deconstructed into components (e.g., sensory,
motor, and cognitive features). By examining intrusive events across diagnostic
categories, common basic biobehavioral processes may be revealed which, in turn,
could facilitate the study of neural processes underlying the behaviors. A model
of cognitive and emotional decision making is presented to provide a basis for
understanding and studying intrusive events. Examples of how the model might
account for the “failure modes” in intrusive events are used to formulate testable
hypotheses, and future interventions that combine multiple treatment modalities are
considered. The chapter concludes with a discussion of the broader cultural context
of intrusive events.
Group photos (top left to bottom right) Judson Brewer, Harriet de Wit, Aurelio
Cortese, Martin Paulus, Emily Holmes, Judson Brewer, Damiaan Denys, Colleen
Hanlon, Peter Tse, Jens Schwarzbach, Martin Paulus, Colleen Hanlon, Aurelio Cortese,
Harriet de Wit, Judson Brewer, Peter Tse, Damiaan Denys, Martin Paulus, Harriet de
Wit, Jens Schwarzbach, Emily Holmes
348 J. Brewer et al.
Introduction
Defining the Phenomenon
What is an intrusion? To which entity does the concept of an intrusion cor-

respond? Is it a single concept that refers to one and the same mental phe-
nomenon, or are there multiple types that arise under different circumstances
and across multiple mental states? Are intrusions different in animals than in
humans, and do they vary across cultural contexts or even time?
The definition by Clark (2005:4) serves as a starting point for our analysis:
…unwanted, clinically relevant intrusive thoughts, images, or impulses [are] any
distinct, identifiable cognitive event that is unwanted, unintended, and recurrent.
It interrupts the flow of thought, interferes in task performance, is associated
with negative affect, and is difficult to control.
Accordingly, an intrusion can be understood both as a clinical symptom and

as a regular mental phenomenon. In as many as 85% of healthy individuals,
for instance, intrusions have been observed in the form of thoughts, images,
or impulses (Rachman and de Silva 1978), experienced as an individual event
that is undesirable, unintentional, and recurrent. In comparison, intrusions as
a clinical phenomenon may be fundamentally different. Their severity may
increase over time, despite or even because of the interventions undertaken by
the patient or caregiver, gradually consuming increased amounts of a patient’s
time and energy. In addition, pathological intrusions rarely diminish on their
own. Unravelling the principle of “reinforcement” that leads to a snowball ef-
fect poses a challenge for both clinical and neurobiological researchers.
As suggested by Clark (2005) and confirmed in our discussions, future em-
pirical studies need to clarify the boundaries of what constitutes an intrusion
for psychiatric clinical and psychological purposes. While a minimum defini-
tion runs the risk of being too restrictive, and may not describe all types of in-
trusive phenomena, it permits us to distinguish between normal and abnormal
intrusions. This, in turn, is needed by diagnosticians to distinguish between
sick and healthy mental events and aid in treatment decision making. As a
minimum definition, we propose the following:
Intrusive events are unwanted, clinically relevant, intrusive thoughts, images, or
impulses that an individual may attempt to resist, but which are out of their control.
A maximum definition, by contrast, needs to incorporate all intrusive phenom-

ena, as exhaustively as possible, into a general descriptive approach. In this
way, an intrusion will be able to be described as a specific mental event or
experience and, regardless of its nature, be studied by different scientific disci-
plines. We propose the following as a maximum definition:
An intrusive event is any interruption in the flow of mental events by an external
(e.g., a ringing telephone) or internal (e.g., a thought) stimulus.
Interventions and Implications 349
Many types of mental events can intrude into consciousness. Verbal and non-
verbal thoughts, mental images, impulses, memories, emotions, desires, and
dreams can all reset the contents of consciousness and be experienced as
unwanted or intrusive. Do all such events have the potential to be clinically
relevant? How does the minimum definition, which primarily encompasses
psychiatric symptoms, relate to the maximum definition, which relates to ev-
eryday experienced phenomena? To what extent do they overlap or deviate
from each other? To what extent are they qualitatively, or perhaps only quanti-
tatively, different from each other? Importantly, if clinically relevant intrusive
thinking takes different forms or domains (e.g., verbal vs. imagery), what are
the implications for treatment? Different treatment approaches may be needed
or optimized for different domains.
In our discussions, we juxtaposed these two definitions next to each other—
one with a psychiatric clinical purpose, the other with a psychological fun-
damental purpose—but wish to emphasize that intermediate viewpoints are
possible. To address our group’s topic, however, we found these artificially
contrasting definitions helpful. A discussion of the philosophical and social
implications of defining the phenomenon is included later in the chapter.
Intrusive Events and Psychiatric Disorders
Across many common psychiatric disorders, intrusive events present as key

symptoms (see Schlagenhauf et al., this volume). Recurrent unwanted thoughts
and images occur in almost every psychiatric disorder and are explicitly de-
scribed among the criteria that must be met for a formal diagnosis in several
disorders. Notably, intrusive events may be problematic symptoms in them-
selves. As such, they are appropriate targets for interventions but have rarely
been identified as a transdiagnostic clinical feature constituting a target for
treatment (cf. Iyadurai et al. 2018). Given their importance across mental dis-
orders, intrusive events may offer insight into the neural mechanisms involved
in the pathophysiology of psychiatric conditions.
In our discussions we de-emphasized intrusive events that are manifest in
perceptual or thought disorders (e.g., hallucinations or intrusive delusions as-
sociated with schizophrenia or psychotic depression) or tics (as in Tourette
syndrome). These may fall into separate categories of events and may notably
lack the “negative affect” component specified by Clark (2005). As a result,
our focus here is on the specific symptoms of intrusive events, not on full psy-
chiatric diagnoses.
Intrusive events vary in their prominence as defining symptoms for differ-
ent diagnoses. Importantly, the pathophysiology of intrusive events may dif-
fer across diagnostic categories and may differ from other symptoms within a
disorder. For some disorders, such as posttraumatic stress disorder (PTSD) and
obsessive-compulsive disorder (OCD), intrusive events appear to be central
defining features and may causally drive other symptoms (see Holmes et al. as
well as Hanlon, this volume). In others, intrusive events may not be the defin-
ing feature of a disorder but rather one of many criteria used to reach a diagno-
sis. For example, craving is a symptom of substance use disorder (SUD), but
SUDs can and do appear without cravings. Similarly, suicidal ideation appears
in major depressive disorder, but the disorder can occur without it. By focusing
on a single symptom, it may be possible to relate the clinical manifestation of
an intrusive event to a dysfunction of neural circuits controlling normal brain
function.
Given the heterogeneous nature of intrusive events associated with different
psychiatric disorders, we recommend that a research program be developed
to examine whether subcategories of intrusive events exist; this information
is needed to establish an association with the underlying neurobiology. More
importantly, intrusive events with a different neurobiological signature may
require distinct treatment approaches. Given the different dimensions of an
intrusive event (e.g., prior experience and expectancies, precipitating events,
temporal sequence of events, emotional breadth, contextual features, psycho-
logical consequences), we offer two clinical vignettes to illustrate the clinical
manifestations of intrusive events in actual patients and demonstrate the inher-
ent complexities.
From an index traumatic event, patients with PTSD typically experience two
or three different, highly vivid intrusive memories both in visual and other
sensory modalities (Grey and Holmes 2008). For example, after a traumatic
road accident, a person might experience vivid intrusive visual mental images
of an oncoming red truck, which originated from the moment just before the
accident. This intrusive image is highly distressing and often associated with
the strong emotions that occurred at the time of the trauma (fear, helpless-
ness). Additional multimodal sensory images may originate from the same
event. For example, the visual memory is also multimodal, comprising sight
of the person’s hand on the driving wheel accompanied by the sound of glass
breaking and the smell of burning. The memory may also be associated with
secondary emotions such as horror and guilt. Well after the event has oc-
curred, an otherwise innocuous visual stimulus, such as a red front door, may
remind the person of the red truck, thus triggering an emotional response.
Emotional states may also serve as triggers for the traumatic event: when a
person experiences a feeling of helplessness about an unrelated event, this
may elicit the feelings of helplessness associated with the traumatic memory.
These experiences may be especially disturbing because the patient has not
associated these external or internal cues to the occurrence of the intrusive
memory, so that they occur as if without warning and as both unpredictable
and uncontrollable. Thus, even though these intrusive memory experiences
(colloquially often referred to as “flashbacks”) may be brief (i.e., seconds,

followed by up to 30 min of emotional response), they can have a powerful
impact on ongoing behaviors and disrupt attention and performance (Holmes
et al. 2017). Intrusive memories not only interfere with normal activities, they
also produce significant emotional distress and physiological arousal which
can be highly disruptive. The experiences typically occur repeatedly, at un-
predictable intervals and in unexpected settings (i.e., elicited by various con-
textual stimuli), and may occur at varying intervals (e.g., one per week to
several per day). Efforts to “push the memory” from one’s mind often fail.
The unpredictable nature of the events can set off a cascade of other symp-
toms, including efforts to avoid reminders of the trauma which may lead to
social withdrawal. A primary goal for treatment is to reduce the frequency and
emotional valence of the intrusive image-based memories.
There are several treatments with a good evidence base for PTSD.
Behavioral treatments provide an interesting example from the perspective of
intrusive memory. Trauma-focused CBT (cognitive behavioral therapy) and
EMDR (eye movement desensitization and reprocessing therapy) both involve
repeated exposure to the trauma memory. For example, trauma-focused CBT,
a particular form of CBT, focuses on the trauma memory but not on verbal
thoughts “about” the trauma (as CBT for depression would). It does this by
using imaginal or in vivo exposure to the trauma memory, which requires the
patient to bring to mind the sensory image-based memory in rich detail. For
instance, over a series of 12 sessions, the patient is encouraged to talk about the
trauma in detail, to “relive” the memory in their mind’s eye, and when possible
to bring in adaptive information for memory updating (e.g., feeling of safety,
that they did not die in the car crash). Initially, this process is typically highly
emotional (and the patient may become upset and cry in reliving sessions), but
it becomes less so over repeated sessions. Patients are taught not to “avoid”
reminders or to push the intrusive memory from their mind. In summary, the
emphasis is on deliberately retrieving the emotional memory in vivid sensory
detail. Over time the memory becomes less vivid, the emotions and meaning
updated, and the number of intrusive memories declines.
EMDR is a related behavioral treatment that requires somewhat less de-
tailed recall of the trauma: when the memory is recalled in a therapy session, it
is done so in the presence of a concurrent task, such as side-to-side eye move-
ments or bilateral beeps. The patient deliberately brings the trauma to mind and
simultaneously performs the task guided by the therapist. Research indicates
that the success of EMDR is related to impact of the concurrent task in work-
ing memory. This is not unlike the brief procedure being developed to reduce
trauma intrusions by using a memory reminder plus Tetris computer game play
(Holmes et al. 2009, 2010; James et al. 2015; Horsch et al. 2017; Iyadurai et
al. 2018; Kessler et al. 2020; see also section below on Future Interventions).
At the end of successful behavioral treatment, the patient should be
able to recall the traumatic event at will, if they wish to, without becoming
overwhelmed. Critically, they will no longer be experiencing frequent invol-

untary, unwanted intrusive imagery-based memories that impair their daily
life. If you have not experienced intrusive events yourself, it may be hard
to imagine how powerful brief intrusive events can be for an individual,
and thus how beneficial it can be to ameliorate them. After trauma, intru-
sive thoughts can carry damaging and toxic meanings for the patient. For
example, a rape victim may have an image of the rapist telling them they are
worthless. Rationally, a victim may know this is not true but still suffer under
extreme distress and shame, brought about by the intrusive events, which
cause them to relive this toxic message vividly. Verbal discussion does not
change the meaning carried by the emotional image (presumably because
it differs neurally), but strategies to change the intrusive event (the image
itself) can, as illustrated here.
It is clinically compelling to see how successful PTSD treatments are cen-
tered on ameliorating the “hub” symptoms of intrusive sensory memories of
the traumatic event. Reducing the emotional efficacy of intrusive events and
the meanings they carry, as well as the frequency of their occurrence, can lead
to substantial improvements in a patient’s quality of life.
Many questions remain: How can models explain the impact of existing
treatments on the reduction in frequency of intrusive memories? How can
treatments be improved to make them briefer and even more effective? Ideally,
we should develop simpler, focused treatments that can help more people glob-
ally (Holmes et al. 2014, 2018).
Obsessive-Compulsive Disorder
In OCD, intrusiveness coincides with the feeling of “being out of control” and
is experienced in different phenomenological domains, as OCD develops over
time. OCD is a process with different clinical stages rather than one single
stage. These stages follow a dialectic interaction in which an intrusive event
elicits a response and the response amplifies the intrusive event. Through dif-
ferent neurobiological adaptations, the course of OCD eventually worsens.
Thus, OCD should be regarded as a disease process that develops through the
amplifying interaction between (the reflection and resistance of) the person
(mind) and the disorder (brain).
Take, for instance, a young mother who recently gave birth to her first child.
She carries an enormous burden of being solely responsible for a helpless and
vulnerable life. Her husband leaves daily for work, leaving her alone at home
with the child. Seeing her young baby in the crib, a thought appears in her
mind: she imagines that she could strangle her baby in the crib and that because
she is alone, no one could prevent her from acting on that thought. In summary:
1. The mere presence of this thought is intrusive because it occurs against
her will. She feels out of control and is unable to volitionally control
her thinking.
2. She is worried because the idea of strangling her baby does not
match the ideal of motherhood she maintains and strives to achieve.
The content of the thought is intrusive because it is not in line with
her identity or expectancy. It is ego-dystonic in that it does not
match her self-image, thereby giving rise to mental discomfort and
distress.
3. She wonders whether she really could strangle her baby. How can she
be certain, given the fact that humans are notoriously unpredictable,
that she won’t in fact destroy that which she most treasures? It seems
that the freedom to commit such a terrible act is in itself so disturbing
that it causes the thought repeatedly to reoccur. In addition, moral im-
plications of the initial thought intrude as well.
4. She feels anxious because the thought confronts her with feelings of
being out of control. The emotional value (anxiety) of the thought is
intrusive. The presence, content, implication, and emotional value of
the thought all have an intrusive quality. Although she actively resists
the thought because it annoys her and feels intrusive, the very process
of reflecting and/or resisting the thought reinforces the frequency and
strength of the intrusive thought. Her efforts to eliminate the thought
may, in fact, enhance its occurrence.
5. The thought becomes obsessional, and her attention is completely
drawn to that one single thought. Obsessionality is a dysfunction of
intentionality: the incapacity to shift focus or attention to another topic,
due to a stronger and longer intentional relation with the mental act.
The thought is intrusive because of its obsessive nature.
6. She cannot suppress the thought; moreover, she is compelled to think
about her obsession. Compulsivity is a dysfunction of sense of agency:
she is forced to think about the intrusion, contrary to her willpower.
The thought is intrusive because of its compulsive nature.
7. Gradually the thought becomes more present and repetitive; it loses its
original meaning, but remains an intrusion because of its duration and
repetition. The thought is intrusive because of its new form or appear-
ance; it is now a full-blown obsession.
8. Obsessions are answered with compulsions. (Note: both obsessions
and compulsions are intrusive, with both having obsessional and com-
pulsive qualities.) Though initially successful in reducing anxiety,
these compulsions gradually become intrusive since the acts have to be
performed compulsively.
9. Eventually, the anticipatory power of the intrusion becomes so over-
whelming that reality testing is disturbed. She does not know anymore
whether she has or has not strangled her baby. Thoughts may become
delusion-like, with psychotic features.
Assessment and Domains of Intrusive Events

Because intrusive thoughts occur in healthy individuals as well as pathological
states, it is important to develop sensitive tools to quantify these experiences.
Healthy individuals experience unwanted intrusive thoughts (e.g., of dirt, con-
tamination, doubt, harm, injury, sex, religion, order, symmetry, superstition)
that are structurally and content-wise similar to clinical obsessions in OCD
patients (Freeston et al. 1991; Langlois et al. 2000a, b) but less severe and
disruptive. Intrusions vary in both the structure and content across individuals
(Clark and Inozu 2014), as well as in terms of frequency, intensity (or distress),
the degree to which the event is being perceived as an intrusion, unexpect-
edness, persistence (duration), controllability, vividness, valence (positive vs.
negative), adhesiveness (durability), and modality (verbal vs. imagery based).
These features provide dimensions that can be used to quantify intrusive events
and to relate them to the underlying neurobiology. Thus, a refined multidimen-
sional quantitative assessment of an intrusive event is critical for developing
quantitative models and to assess the efficacy of interventions.
As reviewed by Clark and Purdon (1995), a number of investigators have
developed questionnaires to assess intrusive events. These questionnaires
distinguish between intrusive events that appear to be triggered by external
stimuli and those that occur spontaneously. In one analysis, it was estimated
that approximately 80% of intrusive events are provoked by an external trig-
ger (e.g., Edwards and Dickerson 1987a). Clark and Inozu (2014) note that
whereas intrusive events in nonclinical samples are context dependent (i.e.,
have external precipitants), clinical obsessions in patient populations appear to
be more spontaneous. They also claim that avoidance of triggers in patients is
associated with more adverse impact of both intrusive events and obsessions.
Clark and Radomsky (2014) identify several aspects that need to be considered
in assessing intrusive events:
• Content and process characteristics (the degree to which the event is
unwanted)
• Discriminant validity
• Relationship to measures of worry and/or rumination
• Degree of self-relevance
• Appraisal variables (e.g., controllability, unacceptability, discomfort,
guilt, dismissibility, unpleasantness)
• Degree of personal responsibility
Investigations of intrusions have taken several approaches, including question-
naires, diaries, and procedures that assess the impact of intentional mental con-
trol on unwanted intrusive thoughts. Each approach has limitations. The most
common approach has been self-report questionnaires:
• The Experience of Intrusions Scale is a five-item measure that assesses
the frequency, unpredictability, and unwantedness of intrusive thoughts,
as well as the interference and distress caused by the intrusions, each on

a five-point Likert-type scale (Salters-Pedneault et al. 2009).
• The Interpretation of Intrusions Inventory consists of 31 items that re-
fer to interpretations of intrusions that have occurred recently. Three
of the above domains are represented: importance of thoughts, con-
trol of thoughts, and responsibility (Obsessive Compulsive Cognitions
Working Group 2001).
• The Obsessive Beliefs Questionnaire focuses primarily on obsessive
intrusion (Obsessive Compulsive Cognitions Working Group 2003).
• The Obsessive Intrusions Inventory consists of a 52-item self-report
instrument designed to assess intrusive thoughts, images, and impulses
that are similar to the aggressive, sexual, and disease-related thinking
characteristic of clinical obsessions (Purdon and Clark 1993).
• The Cognitive Intrusion Questionnaire examines the following do-
mains: frequency, duration, percentage of verbal and image content,
interference, ego-dystonic nature, stimuli awareness, and associated
emotions (Langlois et al. 2000a).
• The Cognitive Intrusions Questionnaire–Transdiagnostic Version items
are grouped based on theoretical criteria into categories labeled intru-
siveness, appraisals, emotions, and strategies, which were selected as
components of a model that encompasses the different ways in which
intrusive thoughts are processed (Romero-Sanchiz et al. 2017).
Whereas each of these questionnaires has a place in a specific context, they are
limited as measures of intrusive events across diagnostic categories. Further,
they were not designed with the goal of investigating the neurobiological ba-
sis of intrusive events. Some of the questionnaires include a broad range of
negative thought content (e.g., anxiety and depressive thoughts) which are not,
strictly speaking, intrusive events. Questionnaires are also limited because in-
trusions are often idiosyncratic and triggered by external cues that are difficult
to describe: their dependence on retrospective self-report of unwanted intru-
sions may not be fully reliable. The questionnaires have not been used in the
context of transdiagnostic, dimensional psychopathology, which are necessary
to determine the heterogeneity of these events. In addition, they have not been
designed to connect to the underlying neuroscience (i.e., the use of domains
that can be mapped to specific brain systems), and the item response char-
acteristics have not been rigorously assessed. Thus, future item bank-based
questionnaires might substantially reduce subject burden by using an adaptive
measurement framework similar to the PROMIS system (Cella et al. 2010).
An alternative to questionnaire-based approaches is the use of patient dia-
ries. Clinicians using behavioral therapies for intrusive events typically ask pa-
tients to monitor their intrusions in a diary (Grey and Holmes 2008). In experi-
mental studies with healthy volunteers, a range of diary measures for intrusive
events have been developed so that participants can report on their intrusive
experience in daily life (James et al. 2016c; Lau-Zhu et al. 2019). Recording
modes can range from pen and paper, to an SMS or an online interface. Diaries
have the merits of being able to be precisely tailored to the research question in
mind; they are sampled in real time and are thus less prone to memory biases
inherent in retrospective self-report. One could imagine that apps will become
useful in this regard.
Although intrusive events may be a component of worry, simultane-
ous administration of worry and intrusive event questionnaires yielded dif-
ferent factor structures. Nevertheless, the factor structure for the strategies
used to counter the thoughts were highly similar for both types of thought.
Furthermore, regression analysis identified interesting relationships between
the strategies, the thought characteristics, and appraisal of intrusive thoughts
and worry (Langlois et al. 2000a, b).
It appears that concern about the personal meaning of the thought is a unique
dimension for obsessive intrusive thoughts (Clark and Claybourn 1997). There
is, however, fundamental disagreement as to whether clinical and nonclinical
intrusive events can be conceptualized along a continuum. Some researchers
focus on thought content: Belloch et al. (2004) found that the ten most fre-
quently occurring thoughts were related to accident, harm, sex, and aggression.
Others focus on the process characteristics, thus emphasizing the intrusive
aspect of the thought (Rachman and de Silva 1978). Interestingly, Lee et al.
(2005) found that the most upsetting intrusive thought is often autogenous; that
is, such intrusions come abruptly into consciousness without identifiable evok-
ing stimuli and are perceived as ego-dystonic, aversive enough to be repelled,
and include sexual, aggressive, and immoral thoughts or impulses.
Intrusive thoughts are thought to be closely related to dysfunctional be-
liefs (Obsessive Compulsive Cognitions Working Group 2003). Thus, there is
an urgent need to assess underlying belief systems associated with intrusive
events. These include:
• Over-importance of thought: beliefs that the mere occurrence of an in-
trusive thought marks its significance.
• Need to control thoughts: beliefs that one can and should exercise com-
plete control over unwanted intrusive thoughts, images, and impulses.
• Perfectionism: beliefs that a perfect response or solution to every prob-
lem is necessary and that even a minor mistake can lead to serious
consequences.
• Inflated responsibility: beliefs that one is liable for causing and/or pre-
venting significant negative outcomes for self or others.
• Overestimated threat: beliefs involving exaggerated estimates of the
probability and/or severity of harm to self or others.
• Intolerance of uncertainty: beliefs that it is necessary to be certain and
that unpredictability and ambiguity should be minimized as much as
possible.
Another key aspect of intrusive thoughts relates to an appraisal, or how the

intrusive events relate to important goals, values, and concerns. Appraisals fol-
low intrusive thoughts and have been linked to future increased frequency.
Appraisal can affect the ability to dismiss the thought, guilt, uncontrollability,
and belief that the thought could come true. It could also affect responsibility,
perceived consequences of the thought, beliefs about the importance of the
thought, and worry that the thought may reflect something about one’s person-
ality (Berry and Laskey 2012). Moreover, in response to intrusive thoughts,
individuals frequently engage in the following strategies:
• Reasoning that focuses on the thought being irrational or unimportant
• Thought replacement geared to distract or stop the worrying
• Social support, talking through, reassurance seeking, physical action,
or doing nothing
While the themes of intrusive thinking are similar across clinical and non-
clinical populations, the appraisal strategies in clinical populations seem
to focus on responsibility and subsequent avoidance strategies. Results in-
dicate that the more distressing a thought was perceived to be, the more
likely participants were to recommend unhelpful strategies (Bomyea and
Lang 2016). Conversely, the less distressing an intrusive thought was, the
more likely participants were to recommend helpful strategies (Levine and
Warman 2016).
Consistent with the appraisal model, Purdon and Clark (1994) found that
the belief that one could act on the intrusive thought and a perceived uncon-
trollability of the thought were important predictors of the frequency or persis-
tence of the distressing intrusion. Freeston et al. (1991) note three distinctive,
dominant response styles: (a) no-effort response (26%), (b) attentive thinking
(34%), and (c) escape or avoidance (40%). Moreover, they found that intru-
sions eliciting escape-avoidance strategies were evaluated more disapprov-
ingly than thoughts eliciting attentive thinking. Rachman (2014) pointed out,
however, that the following issues still need to be addressed:
• Additional information is needed on prevalence in clinical/nonclini-
cal samples.
• The variable content of intrusions needs to be examined as a function
of culture and environment.
• The nature and effect of repugnant versus nonrepugnant intrusions re-
quire examination.
• Further research is necessary on intrusive images and other percepts.
• Experimental investigations of intrusions are needed.
• Randomized clinical trials are needed to examine the effect of different
treatments on intrusions.
Taken together, it will be important to assess appraisal domains and to quantify
response styles associated with intrusive events.
Following work and suggestions by others (summarized above), one ap-

proach to measure and quantify intrusive events is to decompose the experi-
ence into different domains or dimensions. For example, an intrusive event
brought on by a ruminative thought that refers to a past experience associ-
ated with a strong negative emotion could be quantified along the following
dimensions:
1. The degree to which the intrusive event (e.g., a short utterance or an
elaborate verbal instruction) is characterized by a verbal thought
2. The degree to which the intrusive event (e.g., a vivid image, smell,
sound, or other internal perceptual experience) is characterized by an
internal sensory representation in the absence of a percept
3. The degree to which the intrusive event is associated with a positively
or negatively valenced affect (e.g., severe anxiety, guilt, or shame), and
the quality of that affect
4. The degree to which the intrusive event refers to an experience in the
past, present, or is focused on possible future events
This approach would allow us to characterize the degree to which an intrusive
event recalls a distant or recent past event or refers to an immediate or remote
future event. For instance, the intrusive image experienced by an individual
with PTSD might be characterized as relatively low intensity on Pt. 1, high
intensity on the sensory representation (Pt. 2), with an association of guilt
or shame (Pt. 3), and related to the past (Pt. 4). In comparison, an intrusive
thought (worry) in an individual with generalized anxiety disorder might rate
high intensity on Pt. 1, low intensity on Pt. 2, with a focus on anxiety (Pt. 3),
and a primary focus on the future (Pt. 4). Such a decomposition could be used
in large-scale surveys to begin to delineate the frequency of the phenomenol-
ogy of intrusive events and the association of these events with a particular dis-
order. Subsequent statistical analyses (e.g., latent variable analyses and cluster
analyses) could then be used to develop an empirically derived taxonomy of
intrusive events. This approach would permit the severity of the intrusive event
to be quantified and could be an outcome measure for the success of interven-
tion studies.
Finally, this type of decomposition lends itself as a covariate for neuroim-
aging studies to delineate the neural circuitry associated with intrusive events.
For instance, it would be extremely interesting to determine whether individu-
als who have experienced predominantly visual sensory intrusive events show
changes in activation patterns in the visual processing stream, such as the oc-
cipital cortex, compared to individuals with similarly intense intrusive verbal
thoughts hypothesized to show primary changes in the left ventrolateral pre-
frontal cortex.
How could these domains be neurologically implemented? In Appendix
17.1, we present a model that decomposes intrusive events into domains and
then considers how these domains might be instantiated in the brain.
Treatment Approaches
To date, most existing interventions consist of treatments that address each

psychiatric disorder as a whole, rather than intrusive events. One notable
exception is an innovative behavioral treatment for bipolar disorder, which
specifically targets mental imagery-based intrusions through imagery-based
cognitive therapy (Holmes et al. 2019). This same treatment has shown promise
in reducing overall bipolar mood instability (Di Simplicio et al. 2016; Holmes
et al. 2016a). In addition, a preventive intervention has been developed for
PTSD that directly targets intrusive image-based memories of an experienced
trauma (Iyadurai et al. 2018). The development of treatments follows three
primary modalities:
First, pharmacological treatments exist for most psychiatric disorders that
involve intrusive thoughts, although their efficacy varies widely across indi-
viduals and stages of the disease process. Selective serotonin reuptake inhibi-
tors (SSRIs) are the first line of treatment for a broad range of disorders, from
OCD to eating disorders, major depressive disorder, and PTSD. Patients who
do not respond to SSRIs may be treated with serotonin and norepinephrine
reuptake inhibitors (SNRIs), mood stabilizers, or anticonvulsants. Beyond
these classes of drugs, there are numerous experimental treatments (e.g., hal-
lucinogens, ketamine, atypical antipsychotics, and opioid drugs). We note
that both SSRIs and hallucinogens, such as psilocybin, DMT and LSD, op-
erate primarily on the serotinergic pathway: SSRIs increase the presence of
serotonin in the synapse, whereas hallucinogens bind to the 5HT-2a receptor
as a serotonin agonist. In both cases, a postsynaptic neuron will respond as
if more serotonin is present. Some neural network modeling has suggested
that an increase of serotonin in a network can facilitate the dynamic firing
patterns of that network from getting stuck in local minima of processing
behavior. There is a growing popular movement, albeit with limited empiri-
cal evidence, of using low doses (i.e., microdosing) of psilocybin and other
hallucinogens to treat mood disorders, as well as high doses of psilocybin
and ayahuasca to treat depression (Pollan 2018). For SUDs, a different class
of pharmacological treatments exists: drugs used to treat SUDs are typically
specific to the class of abused drug. For example, transdermal or oral nico-
tine may be used for smoking cessation, and opioid partial agonists are used
for opioid disorder. Generally, these drugs take the form of agonists, partial
agonists, or antagonists that target the receptor system where the drug acts.
The idea is that the drug alters the receptor function so that the drug-related
stimulus loses its incentive value. The intrusive event in SUDs often takes
the form of strong cravings or urges to use a drug. Thus, drugs under devel-
opment to treat substance abuse address these cravings as well as the actual
drug use.
Second, behavioral interventions include a broad range of procedures:
CBT, whose techniques include exposure to salient stimuli and reminders,
cognitive restructuring, as well as mindfulness and contingency management

for SUDs, among others. Typical treatment targets are (a) to lessen reactivity
to triggers that initiate the intrusive event or to update the underlying memory
(e.g., after trauma) and (b) to develop strategies to control the user’s response
to the intrusive event. Treatments used to lessen the initial impact and/or up-
date the memory include exposure therapies or repeated experiences with the
thoughts or images in a safe environment. This may be seen as akin to extinc-
tion procedures in operant conditioning. A new form of treatment technique is
being developed by Holmes and colleagues that targets the memory aspect of
an intrusive event by interfering with consolidation or reconsolidation of the
memory of a traumatic event (Holmes et al. 2009, 2010; James et al. 2015;
Horsch et al. 2017; Iyadurai et al. 2018; Kessler et al. 2020). Treatments used
to control the responses include cognitively reframing the meaning of the
eliciting event, developing competing behavioral strategies that are incom-
patible with the immediate response, utilizing social support to minimize the
emotional response, or learning to decrease emotional reactivity. For exam-
ple, mindfulness training has been shown specifically to decrease habitual
reactivity to intrusive events (e.g., craving for cigarettes and food) and has led
in some cases to significant reductions in unwanted behaviors: five times the
quit rates in smoking cessation and 40% reduction in craving-related eating
(Elwafi et al. 2013; Brewer and Pbert 2015; Brewer et al. 2018; Garrison et al.
2018; Mason et al. 2018).
Third, nonpharmacological brain interventions include neuromodula-
tion such as transcranial magnetic stimulation (TMS), deep brain stimulation
(DBS), lesions to specific brain areas (e.g., anterior cingulate cortex), and elec-
troconvulsive shock. These techniques have been used under limited circum-
stances and in highly selected patient populations. DBS, for example, is used
only in severe cases of OCD (Tyagi et al. 2019). DBS targeting the ventral limb
of the internal capsule and the nucleus accumbens is an effective treatment
strategy for treatment-refractory OCD. TMS is approved to treat depression
(Mutz et al. 2018) and is under study for the treatment of OCD (Rapinesi et
al. 2019), Tourette syndrome, PTSD (Kozel et al. 2019), and SUDs (Zhang et
al. 2019).
In many cases, pharmacological, behavioral, and neuromodulatory mo-
dalities are used in combination in various forms. Drugs are used to facilitate
the psychotherapeutic process, such as MDMA and psychotherapy for PTSD
(Mithoefer et al. 2019), whereas behavioral treatments are used in combi-
nation with nicotine replacement therapy for smoking. These different treat-
ments may occur simultaneously or sequentially. In one case, the FDA has
approved a treatment for depression using all three modalities: TMS, SSRI,
and behavioral treatment. As noted, we have no information about the ef-
fects of these treatments specifically on the frequency or severity of intrusive
events, separate from other symptoms of the disorder, and future research is
needed here.
Future Interventions for Psychiatric Disorders
As discussed, very few treatments exist that explicitly target intrusive events
as an aspect of a disease, and for those that do exist, it is not clear how they
actually modify intrusive events. Although intrusive events are included in the
DSM-5 criteria for many disorders, they have not yet been targeted as a distinct
treatment domain.
Treatment innovation is essential and may benefit from being mechanisti-
cally driven and by combining treatment modalities, such as combining phar-
macology with a psychological/behavioral approach (Holmes et al. 2018).
Before going into detail about specific treatment modalities, we discuss how
interventions that target intrusive events might be used to modify core aspects
of an intrusive event: gating, error correction, salience, and evaluation.
Gating Errors
Within psychiatric diseases, intrusive events may result from too much infor-
mation being allowed into awareness, either in a global sense or by a selective
memory gaining access to awareness through a selective gate. This broad con-
struct could be applied to many psychiatric diseases. Intrusive verbal thoughts
associated with general anxiety disorder (see earlier discussion) may be due to
a “weak” gate that allows a lot of information in and ultimately leads to stress
and anxiety. Intrusive image-based thoughts associated with PTSD may be due
to a faulty assignment of salience and/or evaluation that emerges for the trau-
matic event, which allows them to enter into or persist in awareness more eas-
ily than nontrauma-related thoughts. Intrusive thoughts of drug cues in SUD
may be construed as a combination of faulty gating and salience, compounded
by reinforcement learning.
Error Signal/Detection
Intrusive events may also result from a heightened error signal during oth-
erwise normal thought. To achieve ordinary activities of daily living, mental
processes must stay on course and not be derailed by irrelevant streams of in-
formation that constantly come into our brain. A lot of information is processed
on a subconscious level, wherein only the stimuli that are most different from
our expectations are processed consciously. For example, when walking down
the street, an individual is able to maintain posture and balance, typically via
subconscious processing that has become highly automatized. Only when an
event occurs that violates expectations (e.g., when we stumble or suddenly see
an unexpected object in our path) is the walking process brought to the level of
consciousness via bottom-up mechanisms. Alternatively, our own cascade of
ongoing neural processing related to other aspects of life, which are not directly
associated with the goal, can also interrupt automatic behaviors, as when we
suddenly remember that we were supposed to pick up the kids at school. Thus,
bottom-up and top-down internal interrupts exist as well as external interrupts.
The difference between our expectations of internal and external stimuli and
the actual perception of these stimuli can be considered a prediction error.
It is possible, for example, that verbal or visual intrusions in PTSD may
be related to a momentary yet large prediction error. That is, at times, the per-
ceived difference between the thoughts associated with the trauma and the
typical subconscious narrative that is ongoing in an individual’s mind is large
(high prediction error) and triggers the individual to switch their attention to
those momentary thoughts. This is a particularly attractive hypothesis given
the sparse temporal nature of intrusive events in these patients. In daily living,
for instance, many individuals with PTSD are able to function relatively well
in between intrusive events of their trauma. They are often able to work and
care for families, and to conduct their lives for some days without having an
intrusive event. The frequency of occurrence of an intrusive memory of trauma
for some patients may be infrequent (e.g., once a fortnight) whereas for oth-
ers with more severe levels of PTSD, it may be up to every hour. Typically,
measurement tools of PTSD capture this rate of occurrence, although more
research is needed.
Here we address emerging ideas for future treatment innovation based on
the model proposed in the Appendix. Assuming that these reward prediction
errors are coded in a specific neural network, a technology capable of sensing
the magnitude of the violation between expectations and actual input may be
able to change activity in this network in a dynamic manner, pushing it back
into the intended state. This type of closed-loop neuromodulation is currently
used in the treatment of epilepsy. Briefly, sensing and stimulating electrodes
are placed in the brain of individuals with intractable epilepsy in the vicinity
of the seizure focus. The device has a certain “tolerance” for background vari-
ability in the neural activity in the vicinity of the electrode. Once a critical level
of variance is detected (high variance), the device is able to stimulate the brain
and push it back into a healthy state (low variance), thus avoiding the cascade
of a seizure (for further discussion, see Hanlon and McTeague, this volume). It
is easy to see how a similar approach could be used to abolish intrusive events
in a dynamic manner in individuals with PTSD. In veterans with PTSD, for ex-
ample, a device could be trained to identify the “background” levels of activity
present in normal life as well as activity associated with intrusive war-related
memories, and permit the device to push the system back into the healthy range
of error. One problem with this approach, however, is that large prediction
errors are a crucial element for flexible human behavior. So, an autonomous
closed-loop system would have to be able to detect differences specific to the
trauma memory. While this seems like a tall order, because these intrusive
events are so debilitating to the patients, it is reasonable to imagine that the am-
plitude of their prediction error is so large that the device would be able to have
a very high tolerance threshold, and thus only produce stimulation during the
most extreme examples of prediction errors. Alternatively, individuals could

be trained to self-administer the stimulation when they start to become aware
of an intrusive event. More basic neuroscience research with a fast temporal
sampling profile (e.g., EEG, MEG, in vivo recording) is necessary to evaluate
if prediction error is a viable treatment.
Evaluation
One of the most widely implicated behavioral domains thought to be re-

sponsible for the maintenance of intrusive events is an aberrant evaluation
system. Evaluation brings together goal hierarchy (including the current
task, homeostatic goals, etc.), salience, affective, and reinforcement prop-
erties of previously learned behaviors to determine whether to stay on a
certain task or switch. In terms of intrusive events, which can be seen as an
alternate task, evaluation would help to determine whether to switch, and
for how long, to the new task. Clinically relevant downstream manifesta-
tions of the evaluation system include frequency and duration of intrusive
events in consciousness. In light of emerging understanding with regard to
intrusive events, these manifestations can now be linked to specific neural
systems, informed by what is known with current treatment paradigms. For
example, treatments such as CBT target the cognitive elements of intru-
sive events, theoretically changing belief systems related to intrusive events
(e.g., “Is this true?”), whereas mindfulness training targets the relational
affective component of the intrusive event (e.g., “How caught up am I?”).
These examples provide concrete treatment modalities that can be more
critically studied with regard to efficacy. For example, does mindfulness
training (as a modality to improve evaluative accuracy) change the “sticki-
ness” of an intrusive event, and does this change the frequency, duration,
and salience of the intrusive event in the future (see Brewer et al. 2013,
2018, 2019)? Further, the neural systems that are affected by these interven-
tions can be more specifically studied. For instance, recent work with mind-
fulness training has linked reduction in default mode network activity with
a reduction in cigarette smoking (Janes et al. 2019), yet the frequency and
duration of intrusive events related to smoking have not yet been evaluated.
Future studies can use experience sampling or other modalities to determine
if intrusive events reduce in either of these domains.
A Neuroscience-Derived Neuromodulatory Approach to Treatment
Pharmacological treatments may change the sensitivity of the gate that al-
lows items to enter into conscious awareness. For example, benzodiazepines
may blunt emotional response to negative thoughts in a nonselective manner.
Alternatively, there may be potential for combination treatments that couple
a behavioral approach (e.g., one that evokes intrusive events and thoughts,
such as cue exposure), to open the gate, with a pharmacologic agent, to blunt
the response.
Imagine if one could simply “zap” the brain and magically remove the in-
trusive event causing the distress. Such an idyllic scenario is (unfortunately)
still far away, but promising research is moving us closer. A whole new field of
treatment research has emerged around specific brain stimulation techniques
due to their noninvasive nature and simple theoretical motivation. The central
idea behind brain stimulation and modulation techniques takes a systems per-
spective, which considers the brain as the source of any behavior (the outcome
of brain processes). Behavior is therefore the dependent variable whereas brain
circuits, activity, or neurons are the independent variables. If we are to change
a pathological behavior, this systems-level interpretation implies that we have
to find, target, and modulate the brain process or instantiation that is generating
this specific behavior.
Two main approaches have produced important results in recent years:
TMS and neuroimaging-based neurofeedback. TMS is a technique that gen-
erates weak electrical currents within the brain through direct application of
electromagnetic flux over the scalp, noninvasively into brain tissue (see also
Hanlon and McTeague, this volume). At the electrophysiological level, TMS
induces changes in neuronal excitability, which can cause changes in behavior.
TMS has been recently approved by the FDA for use in the treatment of sev-
eral psychiatric disorders, such as addiction and OCD. Future studies need to
evaluate how TMS in combination with other therapeutic forms might be used
specifically to resolve intrusive events.
In neuroimaging-based neurofeedback, “neurofeedback” is defined as a
closed-loop procedure whereby online feedback of ongoing neural activity is
given to the participant for the purpose of self-regulation. Noninvasive neuro-
feedback can be implemented with several neuroimaging modalities. Some of
the most promising proof-of-concept cases derive from functional magnetic
resonance imaging (fMRI) and electroencephalography (EEG)-based neuro-
feedback (Sitaram et al. 2016; Taschereau-Dumouchel et al. 2018a; Keynan et
al. 2019). We will first consider fMRI, due to its potential as an acute interven-
tion (used once or a few times).
With fMRI, one simple option is to use the overall signal strength in one
predefined brain region. This may work well when the target is general (e.g.,
motor action initiation vs. no motor activity); in other cases, however, one
needs to access a specific representation, which by definition would not be re-
trieved through the overall activation level. Therefore, rather than focusing on
the overall activity level of a region (unspecific), machine-learning algorithms
now allow us to infer the precise activity pattern that corresponds to a unique
stimulus, object, or category (specific). This approach, borrowed from machine
learning, is termed multivoxel pattern analysis (MVPA) (Kamitani and Tong
2005; Norman et al. 2006). Going even further, methods have now reached
the point where we can infer the pattern of brain activity in a target participant
from brain activity patterns in surrogate participants (Haxby et al. 2011).
Although these advances have often remained outside the realm of clini-
cal applications, recent innovative efforts have been directed toward utiliz-
ing MVPA in the field of neuropsychiatric disorders. While mainly centered
around the development of markers for certain disorders or their subtypes, with
some remarkable results (e.g., Yahata et al. 2016; Etkin et al. 2019), some stud-
ies have explicitly targeted the prediction of intrusive events from fMRI activ-
ity patterns (see Holmes et al., this volume; Clark et al. 2014b, 2016).
Bringing MVPA to a real-time setting leads us to the concept of neurofeed-
back, which essentially monitors brain activity patterns over time while using
the machine-learning prediction as neuromodulatory input (e.g., to provide re-
wards). If an algorithm is able to detect intrusive events, the system may then
provide a reward to remodel the association between brain state and appraisal
or some other form of modulation to disrupt the cascade of neural events fol-
lowing the emergence of the intrusive event.
In terms of clinical development, due to the intrinsic nature of MVPA-based
designs, neurofeedback interventions can be easily used in a double-blind,
placebo-controlled way. Since we can find activation patterns that correspond
to specific, distinct mental representations, we can also let the algorithm ran-
domly choose which representation will be the target and which the control.
Neither the experimenter nor the patient has to be aware of the category for the
software procedure to be deployed effectively.
Research has shown that neurofeedback may work via reward processing,
learning, and control networks (Sitaram et al. 2016), and that it depends on
reinforcement learning processes (Shibata et al. 2018). During a typical ex-
periment, the machine-learning algorithm monitors activity in a selected brain
region and, at predefined time intervals, computes a (monetary) score reflect-
ing the likelihood that the current activity pattern resembles a template, target
mental representation. Over time, the brain learns to associate the occurrence
of such representation with rewards.
In two proof-of-concept studies, this technique was used to reduce fear
responses in healthy individuals conditioned toward simple visual stimuli
(Koizumi et al. 2016) as well as in participants with subclinical phobia toward
their feared object (Taschereau-Dumouchel et al. 2018a). The physiological
fear responses (amygdala reactivity and skin conductance) were diminished
only for the targeted representation, while a control stimulus elicited unchanged
fear responses. More recently, the same group reported on the feasibility of us-
ing decoded neurofeedback as a target treatment in PTSD (Chiba et al. 2019).
Importantly, throughout these studies, participants were not told about the link
between their brain activity and the amount of reward they received on a trial-
by-trial basis. When asked to make a forced choice about the target and control
categories, participants answered randomly (Shibata et al. 2018).
The aspects introduced so far raise one crucial point: the entire neurofeed-
back intervention can be applied without ever mentioning or displaying the
upsetting content or event to the patient. The enormous advantage that this
could bring is evident: reducing implicit, physiological fear responses could
create the foundation for subsequent, and more effective, behavioral or cog-
nitive therapies. Neurofeedback could then, as a second step, be applied for
maximal effects and learning.
Alternatively, neurofeedback modulation may provide the basis for damp-
ening a physiological, autonomic reaction, which may then facilitate subse-
quent cognitive or behavioral therapies. Current work on fMRI, MVPA-based
neurofeedback can be expanded to include the notion of dynamic brain state.
Indeed, utilizing a very simple correlation of activity fluctuation between two
brain areas, or a pattern of activity representing an object or category, may
be too simplistic, in particular if we are to target states that are more global
(e.g., attention, emotions, arousal, interoception). Spatiotemporal oscillatory
patterns describe relatively well these global states, distinguishing between
rest and task-based efforts (Vidaurre et al. 2017; Bolton et al. 2018).
Rather than acting on the intrusive events themselves, dynamics-based
neurofeedback interventions could target and modulate an overall affective or
cognitive state. Here, the rationale is different: prepare the brain to receive the
treatments targeting specific aspects of the pathology in the individual.
fMRI neurofeedback carries high costs, immobility, and the requirement for
specialized operating personnel. This greatly hampers the scalability of this ap-
proach, particularly if we think about going beyond acute treatments.
As such, EEG-based neurofeedback holds potential from a different per-
spective. New generation EEG headsets have relatively low costs, are small
in size, and are almost “plug-and-play” ready for use. We can envisage EEG
products in the near future that could be used autonomously by patients on a
daily basis and essentially without many constraints in terms of location or
function, allowing for real scalability. Proof-of-concept EEG neurofeedback
has been demonstrated with specific mental states while targeting deep brain
structures, such as the posterior cingulate cortex which is part of the default
mode network (van Lutterveld et al. 2017), training stress resilience through
electrical fingerprint (Keynan et al. 2019), as well as reduction, consolidation,
and personalization of lead placement (Pal et al. 2019).
A different form of feedback might therefore involve generating an exter-
nal interrupt when an undesirable thought pattern has emerged. For example,
MVPA of EEG signals might allow ruminative thoughts associated with de-
pression to be decoded. If detected, an EEG headset could then be designed
that might vibrate or emit a tone whenever rumination had exceeded, say, ten
seconds, bringing the patient back to the present.
Combination Treatments
A significant future direction in the treatment of intrusive events is the possibil-

ity of combining treatment modalities. The interventions used in the treatment
of intrusive events fall broadly into three categories: behavioral, pharmacolog-
ical, and neuromodulatory. Some of these treatments have already been com-
bined. For example, MDMA is currently under investigation for the treatment
of PTSD, but a central feature of the treatment is that the drug is administered
in the context of psychotherapy. Other examples include the use of behavioral
treatments combined with nicotine replacement therapy for smoking cessa-
tion or other pharmacological-behavioral combinations for treatment of other
addictions.
An Innovative Approach to Targeting Intrusive

Events in Trauma-Related Disorders
In PTSD, trauma is induced through vivid, emotionally laden intrusive memories

that take the form of sensory multimodal mental images of specific moments ex-
perienced during the traumatic event. Current treatments are available to reduce
these memories or their consequences (e.g., trauma-focused CBT), but they can
be hard to implement on a large scale (to reach more people), and evidence-
based preventive interventions after trauma are lacking. Recently, Holmes and
colleagues developed a brief, noninvasive behavioral treatment that reduces both
the establishment and the maintenance of intrusive memories after a traumatic
event (Iyadurai et al. 2018). It should be noted that the treatment focuses on
specific intrusive memories, rather than the whole disorder of PTSD. The type
of intervention developed may be applied either shortly after the trauma, or later,
well after the memories have been established (Holmes et al. 2009, 2010; James
et al. 2015). The procedure consists of the following steps:
1. Subject is instructed to recollect the event briefly via images and
thoughts.
2. Subject participates in a 15-minute Tetris task that involves mental
rotation.
This simple procedure can be implemented shortly after the actual event or
after a retrieval of the event later in the process. It may be administered in a
hospital emergency room or acute ward within six hours of the traumatic event
(Horsch et al. 2017; Iyadurai et al. 2018). In these studies, the intervention
reduced the number of intrusive memories in the following week by approxi-
mately two-thirds. It can also be used to treat intrusive memories of traumatic
events that occurred many years ago. In one study (Kessler et al. 2018), in-
patients with complex PTSD received the treatment with intrusive memory,
resulting in a reduction in the frequency of those intrusions (compared to non-
targeted intrusions). Further research is required.
This novel treatment approach is derived from cognitive psychology and

experimental psychopathology related to memory processes. Performing a vi-
sually demanding task shortly after encoding or retrieving a remembered men-
tal image, while holding the image in mind, competes for working memory
resources in a way that interferes with consolidation/storage (during the initial
experience) or reconsolidation (after the image is retrieved later). Remarkably,
performing this simple Tetris task immediately after thinking about the image
reduces the frequency of intrusive mental images (Iyadurai et al. 2018). The
technique is based on various assumptions:
• Intrusive memories of trauma comprise sensory mental imagery (Grey
and Holmes 2008).
• Intrusive memories can be altered shortly after an event or at retrieval:
memory consolidation/reconsolidation (Visser et al. 2018).
• The capacity of people’s working memory is limited (Baddeley 2003).
• Visuospatial tasks compete for resources in working memory with a
mental image; that is, those that would be needed to (re)consolidate
intrusive mental images (James et al. 2015).
Thus, engaging in a visuospatial task, such as a highly visually demanding
computer game like Tetris, at a time when the mental image of the intru-
sive event is active may reduce the reoccurrence of distressing images later,
as well as the level of distress and vividness associated with them. Further
research is needed to develop this relatively simple and brief behavioral inter-
vention approach. It lends itself to be studied as part of a combination treat-
ment approach, as discussed above. The difference between this theoretical
approach and many others is the focus on a single symptom (intrusive events)
and a neuroscientific account about the underlying mechanisms bringing
about effects.
Implications Beyond Neuroscience and Psychiatry
Philosophical and Social Implications of Definitions
Although the original definition by Clark (2005) and its minimum and maxi-
mum derivatives are intuitively very recognizable, the appearance of clarity
is misleading. The problems that come to light reflect some important philo-
sophical and social assumptions.
What exactly is meant by unwanted? Is it a mental event that is unwanted
by the individuals themselves? Or does this represent what is unwanted by
members of society, relative to what counts as unacceptable, abnormal, or even
moral or immoral (e.g., expressions of sexuality or aggression)? Is what is un-
wanted not wanted relative to short-term individual goals, such as wanting to
drink water, eat a marshmallow, or court a colleague? Or is what is unwanted
defined relative to interference with the fulfillment of long-term goals, such as

completing a PhD program or maintaining a marriage? Are intrusions perhaps
unwanted in light of the conflicts they present to a person’s ideal image, such
as becoming a good scientist? Certainly the notion of being unwanted depends
on a broader philosophical perspective: What kind of human being do we have
in mind as our ideal? Are mental phenomena and interventions that foster this
ideal more wanted than those that lead away from this ideal? In sum, there are
multiple ways to be unwanted, whether by the individual or society. Should we
privilege one type of being unwanted over another?
Inherent in the definition of intrusive thinking is the notion that a lack of
control over the flow of mental events is unwanted. The implication is that in-
trusive mental events are unwanted at least in part because they are not subject
to volitional control. Overcoming a loss of control, in turn, implies that mental
events could be made subject to volitional control. But what makes a mental or
neural process volitional versus nonvolitional? In the motoric domain of eye
movements, for instance, changes in pupillary size (or nystagmus or microsac-
cades) seem to proceed automatically, regardless of intentions or plans held in
working memory or reportable by a subject. In contrast, saccades and smooth
pursuit eye movements are subject to flexibly updateable plans or intentions
held in working memory, which in turn can be reported by a subject (e.g., “I
was looking around for my child”). Implicit in this is the notion that different
intentions would lead to different eye movements. Similarly, volitional control
of mental events implies that different mental events could and would have
arisen had intentions or plans been different. There seems to be an implicit as-
sumption here that events could have turned out otherwise than they did, had
volitional control of actions or thoughts been different.
This raises the age-old problem of free will. If, under a deterministic world-
view, mental events could not have turned out otherwise, then whether a par-
ticular intrusion would happen at a given time is dictated by the laws of physics
before one was even born. The occurrence of involuntary intrusions seems
consistent with the possibility that they could not have turned out otherwise.
It is the notion that they can be brought under voluntary control that may be
undermined by determinism. But how would indeterminism save the possibil-
ity of voluntary control of mental events? It might seem that randomness is as
little subject to agentic control as events determined to happen before one was
even born. Is there a middle path between the apparent lack of voluntary con-
trol that arises either under determinism or utter indeterminism? If yes, what
might that solution be?
Our definitions of intrusion demand that we specify what it means to be
clinically relevant. No absolute answer is possible, only one relative to a given
clinical approach, of which there are many. Moreover, relevance depends not
only on the clinical picture, but also on the person being treated, as well as the
diagnostician and the social context or culture in which the patient and diag-
nostician find themselves. This implies that it is never possible to determine
with objective certainty what an intrusion means in the psychopathological

sense. It is therefore impossible to find a single neurobiological substrate of an
intrusion without first specifying with great precision what exactly one means
by seemingly basic terms such as “unwanted,” “control,” “volition,” or “clini-
cally relevant.”
This definition of an intrusion refers to the fact that an intrusive mental
event should be distinct or temporally punctate. We may try to apply this defi-
nition to “mind wandering,” which is sometimes considered an intrusive event.
Mind wandering, however, is often difficult to distinguish from regular thought
patterns or normal free associations. It is certainly not punctate, but rather du-
rationally extended. Would we regard the transition to mind wandering as an
intrusion?
Moreover, is an intrusion by definition negative? Not every intrusion is ac-
companied by a negative emotion. Some intrusions, such as love or manic
thoughts, are experienced as pleasurable, or may be concurrently both wanted
and unwanted. Further, some types of intrusion are wanted by the subject, but
viewed as intrusive by others in society. For example, in a traditional Nepali
village, where marriages are typically arranged, the statement “I have fallen in
love” might be met with concern rather than happiness, whereas in the West
this transition to a potentially obsessive mental state is commonly regarded
as positive. Patients with bipolar disorder report experiencing highly positive
mental imagery intrusions associated with their mania, but these may also be
diagnostic for a major psychiatric disorder (Ivins et al. 2014). Interestingly,
one study has demonstrated that positive intrusive thoughts can be induced
even in healthy adults, suggesting that they are open to experimental manipula-
tion and study (Davies et al. 2012).
Although our focus in this chapter has been on the neuroscientific basis
of intrusive events, their clinical relevance, and interventions, consideration
should also be given to the degree to which intrusive events affect our society
or can be understood outside of a solely biological or medical perspective.
Radomsky et al. (2014) found that intrusive events are experienced across a
large variety of cultures. They concluded that there were far more similarities
than differences across different cultural sites and that the contents centered
worldwide on themes of contamination, aggression, doubt, blasphemy, im-
morality, sex, victimization, and miscellaneous intrusions. Culture seems to
influence the content of intrusions to some extent but not its prevalence (Clark
and Inozu 2014). Moreover, the relationship between appraisals, control strate-
gies, and the frequency and distress of intrusive events appears invariant across
countries. This has been confirmed by others, who conclude that there is a cer-
tain degree of universality regarding the prevalence of obsessive, dysmorphic,
hypochondriacal, and eating-related intrusions across a variety of countries
and cultural contexts (Pascual-Vera et al. 2019). Nonetheless, there are cultural
differences in how intrusive events are interpreted. For example, according to
Luhrmann et al. (2015), voices are experienced by schizophrenics in the United
States as primarily negative, in particular, as violations of thinking, whereas in

India and Ghana they tend to be interpreted in a more positive light, involving
relationships with the presumed speakers. Understanding intrusive events in
different cultural contexts will constrain the degree of generalizability of any
neuroscience-based model of intrusive events. Future research should examine
individuals who report similarly frequent or intense types of intrusive events in
different disorders, but who are members of different cultures and have varied
responses to intrusive events. Such an approach is not unlike extracting dif-
ferent computational models from psychotic and nonpsychotic hallucinators
(Powers et al. 2017). Taken together, intrusive events can be a fruitful and
important topic of research that extends the biological sciences and can pro-
vide important information about how society should consider managing these
phenomena in the future.
Of particular interest is the change in the nature of intrusions experienced
by people in our modern societies. Those who came of age in the 1980s or
earlier effectively grew up in an analog world. Since the 1990s, however, with
the advent of personal computers, the Internet, and smartphones, our society
has become digital. Because the last generation to be raised without digital
devices is still alive, now may be an opportune time for this generation, which
has experienced both the analog and the digital world, to reflect on the pros and
cons of this societal and personal transformation.
Modern society is now deeply penetrated by mobile devices, which can be
viewed as “intrusion machines.” Screen media activity (SMA) is ubiquitous
worldwide and among the most salient recreational activities of children and
adolescents. Children and adolescents spend about 40–60% of their time after
school engaged in SMA (Arundell et al. 2016), and nearly 97% of U.S. youth
have at least one electronic item in their bedroom (Hale and Guan 2015). A
heated debate has emerged on whether SMA is associated with psychological
and social problems (Ferguson 2017; Twenge et al. 2017). However, media
behavior is complex, encompassing a variety of activities, such as social and
nonsocial Internet use, gaming, as well video or TV viewing. For example,
whereas males are more likely to engage in video games with a higher poten-
tial for excessive use (Choi et al. 2015), females engage more with social me-
dia (Schou Andreassen et al. 2016) and exhibit more excessive cell phone use.
Moreover, gaming has replaced sedentary screen time, such as TV viewing,
Internet usage, and nonactive gaming (Simons et al. 2012). In the near future,
with 5G technology, individuals may be able to interact with numerous devices
on an almost constant basis. For instance, whereas currently a cell phone might
signal the arrival of a text, email, or message (the intrusive event), we may in
the future be alerted by our refrigerator, car, air-conditioning system, or other
devices that we use in our daily life. Thus, technologically based intrusive
events could have the potential to seriously affect healthy individuals and pos-
sibly to a greater extent any individual who is cognitively or affectively com-
promised (i.e., a person with a psychiatric disorder). Moreover, the constant
engagement with SMA may affect both brain structure and function, and in
turn make individuals more susceptible to the experience of both physiologi-
cal and pathological intrusive events. The investigation of mobile technology
on cognition is still in its infancy (Wilmer et al. 2017), and our knowledge
about the impact of this technology on intrusive events is nonexistent. Thus, an
important goal for future studies should be to determine whether such devices
could contribute to the exacerbation of psychiatric disorders, characterized by
frequent and/or severe intrusive events.
Another perspective on intrusive events and their pathology is the notion
that an individual who experiences a low level of need for control might not
experience intrusive events as problematic or in need of treatment. This raises
the question whether the fact that we are focusing on neuroscience, clinical
consequences, and interventions associated with intrusive events is a by-prod-
uct of our society’s focus on “over control.” Interestingly, individuals with
strong beliefs about controlling thoughts are more likely to experience dis-
tressing intrusions, both with and without meta-awareness, compared to people
with weaker beliefs (Takarangi et al. 2017). Thus, it would be interesting to
examine the frequency, severity, and clinical consequence of intrusive events
in different societies that place different emphases on cognitive control. Along
similar lines, an intrusive event might only be assessed as an unwanted pertur-
bation if the individual has a concept of causes and effects, which is a historical
consequence of the Enlightenment period of the seventeenth and eighteenth
centuries. In prior societies there was a greater emphasis on a teleological
framework within which experiences and events were interpreted through the
lens of their potential function, end, purpose, or goal. In this context, an in-
trusive event may be experienced as something that is necessary to lead to a
particular goal rather than an unwanted distraction. Thus, it may be interesting
to conduct a historical literature analysis that focuses on the characterization of
intrusive events before and after the Enlightenment period.
Let us return now to intrusive events and tie them in with the notion of
frameworks of meaning. If efforts to suppress intrusive events tend to exac-
erbate intrusive events, then perhaps efforts should be made concerning how
best to suppress such suppression, or, on the contrary, how best to facilitate
the expression of intrusive events so that the salient issue so expressed can
be processed in a healthy manner. Other cultures have created modes for the
expression of “forbidden” emotions. Ancient Greek drama often centered on
creating tensions that would then lead to emotional catharsis. Rather than sup-
press unwanted emotions, such as lust for forbidden objects of desire, as in the
play Oedipus Rex, these emotions were vented in a manner that was safe for
society. Even modern European cultures have aspects that are reminiscent of
catharsis. For example, Carnival is a venue for the expression of carnal desires
and behaviors that in other times would be regarded as deviant or dangerous.
How might catharsis be exploited in the context of existing or new therapeu-
tic methods? One possibility would be role-playing, where a traumatic event
or idea is expressed in a manner such that emotions can be safely expressed.

Indeed, by pretending to reenact a particular traumatic event or relationship,
the possibility emerges of having it not only expressed, but of expressing it
in a new way, perhaps having the event turn out differently than it in fact did.
Another avenue might be to build on the tragedies and plays of Ancient Greece
by creating “virtual worlds,” perhaps exploiting movies or virtual reality tech-
nology, where pent-up and suppressed emotions and desires could be released
virtually, rather than through real acts in the life of the patient.
According to this view, intrusive events are analogous to salience signals
arising from exogenous attentional circuitry. Just as the sudden motion of a
tiger demands an interruption of the current plan engaging conscious thought
and planning, so that a new plan (in this case, to escape the tiger) can be gen-
erated, intrusive events are salience signals that enter consciousness because
there is an unresolved issue that requires executive control circuitry to come up
with a plan to resolve the unresolved issue. If this view has validity, then sup-
pressing the salience signal might be about as effective as attempts to suppress
hunger or thirst signals. The reason these signals barge into consciousness is so
that goals can be generated by executive planning areas that will resolve them
(e.g., coming up with a plan to get food or water). If unresolved emotional or
cognitive issues barge into consciousness, a better approach than suppression
may be to find ways to resolve the unresolved issue. Catharsis and role-playing
have already been mentioned. Other possible methods may involve unortho-
dox techniques that are considered quite orthodox in non-Western traditions.
For example, according to Kundalini yoga, emotions are stored in the body
and can be activated with certain bodily actions, such as breathing patterns of
physical exercises. By intentionally invoking the breathing pattern associated
with fear or calmness, say, the mental state that normally accompanies such
breathing patterns can also be invoked and perhaps processed in a manner
subject to volitional control. If unresolved emotional issues are stored in part in
the body, or in bodily patterns of action, perhaps the Western tradition can gain
insights from other traditions concerning the “cleansing” of stored psychologi-
cal tension and pain.
Conclusion
Intrusive events are emerging as an area of interest that can help further our
collective understanding of basic brain function, psychiatric conditions, and
their treatment. Characterizing domains in which intrusive events manifest in
psychiatric disorders may help their characterization, diagnosis, and treatment.
Basic heuristic models can inform how normal brain function can go awry
due to faulty systems, including gating, salience, evaluation, and prediction
error detection. From these models, current and evolving treatments (and po-
tential combinations) can be tested for target engagement, specificity of effect
and efficacy with regard to reduction of frequency, duration, and salience of

intrusive events. Future research would benefit from mechanistically framed
and neuroscience-based approaches with specified targets and tangible target
engagement and clinical outcomes. Our proposed model (see Appendix 17.1)
provides an example of one such emerging approach of a behavioral interven-
tion that specifically targets intrusive events based on theories of memory (re)
consolidation and cognitive task interference.
Appendix 17.1: Proposed Model
Our model begins with a sequence of goals that are represented in the brain
(Figure 17.A1). A mental workspace keeps the present goal in mind, allows
operations to take place over representations held in the workspace (area 4,
working memory), and takes into account prior knowledge, context, and other
system constraints. Salience helps the organism determine how important
a potential interrupt is and interacts with an evaluative system to determine
whether to stay on task or to interrupt it and, if so, which task to then prioritize
as the task to do next. A goal maintenance system helps the mental workspace
maintain the present goal using feedback loops that afford the minimization
of prediction error signals. This may happen via the enhancement of gating of
potential interrupt inputs, or the inhibition of the salience of potential interrupt
signals. The goal maintenance system evaluates deviations from the trajectory
leading toward fulfillment of the goal. Such prediction error signals are used
dynamically and cybernetically to correct the present trajectory to minimize
that error, similar to when a heat-seeking missile alters its path to hit its target.
External interrupts (not internally generated) may orient the organism to
unexpected inputs from the external world (e.g., when we hear a loud sound
or see a sudden motion): if the magnitude of the prediction error (i.e., between
what was expected and what in fact occurs) is large enough, the organism ori-
ents to the external stimulus.
Internal interrupts can be both bottom up and top down. Bottom-up sys-
tems that can generate interrupts include systems that maintain physiological
(e.g., hydration) and nonphysiological (e.g., happiness) goals that are separate
from the current goal. When salient enough, these interrupts provide inputs to
the mental workspace to force a reprioritization of what to do next (stay on
task or switch tasks). Subjectively these signals are experienced as, for ex-
ample, thirst, hunger, lust, or a need for oxygen, salt, or sleep. Other bottom-
up systems include reward/punishment and other evaluative systems but may
not have intrinsic homeostatic functions. These may be experienced as, for
example, fear or other emotions, such as anger. In addition, the memory sys-
tems may automatically retrieve memories which can then appear to “pop” into
consciousness.
Top down 2
internal interrupts: 4
6 Goal 1 4
Default mode intrusive thoughts 4 2
sequencing and
current goal stack 6
network 2
1
2 4
Working memory: 5
Exogenous
attentional Processing present
goal, the perceptual Current goal 5 6
interupts maintenance 5
world and world model,
executing current and switching 6 5 3
systems 5
mental operations 3 3
Lateral 3
External interrupts:
Intrusive disruption
Reward Pain, Gates

Various Various Fear Memory punishment somatic Excitatory inputs
3 desires emotions system systems
systems signals Inhibitory inputs
Bottom up
internal interrupts: Intrusive images, feelings
Figure 17.A1 (1) Neural substrates: frontopolar cortex (e.g., Brodmann area 10). (2) Ventral attentional network: ventrolateral prefrontal cortex,
temporoparietal junction. (3) Numerous subcortical systems including amygdala, hippocampus, nucleus accumbens and ventral tegmental area,
brain stem, and thalamus. (4) Dorsal attentional network: dorsolateral prefrontal cortex, posterior parietal cortex, superior parietal lobule, and
intraparietal sulcus. (5) Cingulo-opercular control network: dorsal anterior cingulate, frontal operculum, basal ganglia; for emotional regulation
ventromedial prefrontal cortex (BA 12) and orbitofrontal cortex (BA 11). (6) Default mode network: anterior medial prefrontal cortex, posterior
375
cingulate, and temporoparietal junction.
Thus, under normal conditions, a goal is maintained using gating, evalu-

ation, and prediction error monitoring and minimization. Under normal con-
ditions, gating is appropriate, evaluation is accurate, and prediction error is
accurate and dynamically corrected. Under pathological conditions, gating
can become too weak or too strong, and prediction error and evaluation can
become inaccurate. This model suggests that any one of these processes, or
a combination of gating, evaluation, and prediction error, can go awry and
thereby cause a pathological condition.
Failure Modes
Pathological conditions may arise when an individual fails to stay on target in

the pursuit of the current goal because
 the prediction error (Figure 17.A1, between areas 4 and 5) does not arise
(e.g., one should feel guilt for a misdeed, but does not) or it is inaccurate,
 evaluation (i.e., signals generated from the different components of Figure
17.A1, area 3) becomes inaccurate, or
 prediction errors do not get eliminated after a course correction takes place.
A system can have a number of modes of failure when one attempts to fulfill
one’s goals (see Table 17.A1). For instance, in a normal state, if the current
goal is to go to a movie, the following thought might arise as a bottom-up
interrupt: “You didn’t study enough for your exam next week!” If gating is
appropriate, that thought need not get into the mental workspace. If the gate
is too strong, appropriate evaluation does not enter the mental workspace
when it should, and one does not study. Once in the mental workspace, if
the thought is deemed to be accurate (“Yes, you did not study enough”), the
system changes its plan; namely, to study instead of going to see the movie.
Consider a case of anxiety, where the same thought arises: “You didn’t
study enough!” If gating is too weak, the thought arises too easily or comes
in more frequently, moving it toward the spectrum of intrusive events. If there
is an excessively strong gate, the thought does not enter consciousness and
one does not feel a need to reevaluate the present plan, leading to the unhappy
result that one does not study when indeed one should. Once in the mental
workspace, if evaluation of the thought is inaccurate (you have indeed studied
Table 17.A1 Modes of failure.
Normal Pathological
Gating Appropriate Too weak, too strong
Salience Appropriate Too little, too much, etc.
Evaluation Accurate (In)accurate
Prediction error Accurate (In)accurate, failure to reset
enough when you think you didn’t), the thought can become pervasive, lead to
excessive worry, or inappropriately change the task to studying.
Salience signals can be faulty when studying seems much more appropri-
ate than necessary, and the task is changed from going to a movie to studying.
Another example of inappropriate salience occurs when one is studying, and
studying appears so much more important than sleep that one continues study-
ing until the moment of the exam, paradoxically hampering performance.
An inaccurate prediction error signal arises when one is on task, but none-
theless gets a signal that one is off task. For example, when someone is study-
ing for a test, one feels anxiety despite engaging in studying, which can para-
doxically undermine test preparation through perseverative worry (instead of
studying).
An additional pathway that can compound mode failures emerges from re-
inforcement learning to combine with the faulty elements described above. If
the threshold of the gate is too low, the interrupt changes the ongoing plan. The
change itself (because it is new) can be reinforcing. This reinforcement leads
to increased salience of the event that produced the change in plan. As salience
iteratively increases, the likelihood increases that the interrupt that is linked
to the new plan is going to disrupt other plans in the future (positive feedback
loop). If the new plan is now in place, the stronger salience of the interrupt sup-
ports its maintenance. For example, with perseverative worry, the worry led to
a change in plan (didn’t go to the movie), which then led the person to study
more and indeed feel better, which then functions as a reward signal that leads
to a reinforcement of perseverative worry in the future. Such a simple mecha-
nism could account for the common finding that OCD worsens with time. An
important question for future research into effective interventions will be how
to rein in this positive feedback loop afforded by reinforcement learning.

08自由意志 Peter W. Kalivas, Martin P. Paulus - Intrusive Thinking - From Molecules to Free Will-MIT Press (2021)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

08自由意志 Peter W. Kalivas, Martin P. Paulus - Intrusive Thinking - From Molecules to Free Will-MIT Press (2021)

Uploaded by

Copyright:

Available Formats

The Day Sue Almost

What Are the Circuits

Mesocorticolimbic Dopamine- and Emotive-

Reward Prediction Errors

In the mid- to late 1990s, computational neuroscientists and computer scien-

the enactment of conditioned responses (Flagel et al. 2011b; Ostlund and

Model-Free and Model-Based Processes

In a simple cue–reward learning task, where there is spatial separation of the

Integration of Appetitive and Aversive Information

With systematic interrogation of dopamine neurons in the ventral tegmental

proposed that appetitive and aversive processing by dopamine neurons is seg-

More Complex Stimulus–Stimulus Associations

multidimensional encoding seems to be especially important to support model-

While the above discussion diﬀerentiates associations that represent specific

Updating and Inhibition of Intrusive Thought

The resistance to extinction of intrusive memories may be especially maladap-

Therapeutic Strategies Using Updating Mechanisms

In addition to providing a potential pathological mechanism underlying the

types of approaches is that diﬀerent types of stimulus–stimulus associations

Intrusive thoughts and actions are highly debilitating symptoms associated

Actions and Habits

It is now well recognized that the performance of instrumental actions (i.e.,

Common Tests of Goal-Directed Action

Loss of Control and Intrusive Action

Some forms of psychopathology that result in intrusive or compulsive actions

we found that schizophrenic subjects had a severe deficit in the sensitivity

The Dysregulation of Goal-Directed Action

It would be interesting to know whether the same kind of dysregulation oc-

selection–evaluation–execution sequence lies at the heart of goal-directed instru-

Cognitive Control of Action

Neural Bases of Pavlovian-Instrumental Transfer

A number of structures have been found to be involved in the way predictive

latter process is controlled by the giant aspiny cholinergic interneurons that

inhibitory response-outcome association will be disinhibited and released as an

Neural Bases of Intrusive Thoughts and Actions

Summary and Conclusions

Corticostriatal circuits play a central role in the induction of intrusive thoughts

What Is the State-of-the-Art

emotional disorders, such as anxiety, depression, or schizophrenia. For many

Observing, Recording, and Manipulation

The architecture of the central and peripheral nervous system is composed of a

Limitations of these new technological recording methods include the

Neuronal Actuators and Related Devices

Electrical recording oﬀers precise, high-resolution information related to spe-

three categories and act to initiate enzymatic activity, dimerize, or change

development of various sensitivities, color activation spectra, and functions

Although optogenetic and chemogenetic approaches have provided unprec-

Detection and Visualization of Neural

In vivo Calcium and Voltage Imaging

Advances in genetically encoded calcium indicators (GECIs) represent the

for simultaneous multicolor imaging of diverse genetically or circuit-specific

Imaging and Detecting Neurotransmitter and Neuromodulator Activity

A very recent and exciting development in biosensors for other biomolecules

Shannon L. Gourley, Antonello Bonci,

Understanding the neurobiology of intrusive thought requires unfettered and

Defining Intrusive Thought in Biological

to capture, operationally and quantifiably, in an experimental subject are

Table 5.1 Validities and considerations in designing research strategies.

Construct validity The interpretability, meaningfulness, or explanatory power

Conceptualizing Intrusive Thought as an

Embedded in the argument that nonhuman experimental systems have utility

ACC Insula ACC Insula

Figure 5.1 Intrusive thought is an emergent property of multiple systems. We propose